<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Maxime Vono on Medium]]></title>
        <description><![CDATA[Stories by Maxime Vono on Medium]]></description>
        <link>https://medium.com/@maxime.vono?source=rss-a90c05078adc------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*FsI3fXR8zUyH1bzcO9yclg.jpeg</url>
            <title>Stories by Maxime Vono on Medium</title>
            <link>https://medium.com/@maxime.vono?source=rss-a90c05078adc------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 08:57:49 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@maxime.vono/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Apple’s Private Ad Measurement]]></title>
            <link>https://medium.com/criteo-engineering/apples-private-ad-measurement-52dcd41021f7?source=rss-a90c05078adc------2</link>
            <guid isPermaLink="false">https://medium.com/p/52dcd41021f7</guid>
            <category><![CDATA[privacy]]></category>
            <category><![CDATA[adtech]]></category>
            <category><![CDATA[criteo]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[privacy-sandbox]]></category>
            <dc:creator><![CDATA[Maxime Vono]]></dc:creator>
            <pubDate>Tue, 09 Jan 2024 07:47:30 GMT</pubDate>
            <atom:updated>2024-01-09T07:47:30.217Z</atom:updated>
            <content:encoded><![CDATA[<h4>And the interplay with other browser vendors’ proposals</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*USxJDzUUZ2gpgVg7sefm0A.jpeg" /><figcaption>Source: <a href="https://unsplash.com/photos/one-red-apple-CoqJGsFVJtM">https://unsplash.com/photos/one-red-apple-CoqJGsFVJtM</a>.</figcaption></figure><h3>Introduction</h3><p><strong>Context.</strong> Criteo stands for a leading AdTech company, notably shaping the future of the industry that funds the web. Indeed, we have been actively working on the Privacy Sandbox, collaborating with Google Chrome for four years, contributing to other browser vendors&#39; proposals via GitHub issues, and writing several <a href="https://medium.com/criteo-engineering/tagged/privacy-sandbox">Medium articles</a>. To cite a few of our contributions, we, for instance, recently published a series of articles focusing on understanding and discussing the technologies underlying current browser vendors’ proposals for <em>attribution and reporting</em> use cases:</p><ul><li><a href="https://medium.com/criteo-engineering/an-introduction-to-pets-for-attribution-and-reporting-7e8cde13eca3">An Introduction to PETs for Attribution and Reporting</a> ;</li><li><a href="https://medium.com/criteo-engineering/pets-in-advertising-scenarios-for-secure-multi-party-computation-4d2e667e5ecd">PETs in Advertising: Scenarios for Secure Multi-Party Computation</a> ;</li><li><a href="https://medium.com/criteo-engineering/pets-in-advertising-scenarios-for-trusted-execution-environments-9d0264c57325">PETs in Advertising: Scenarios for Trusted Execution Environments</a> ;</li><li><a href="https://medium.com/criteo-engineering/pets-for-attribution-and-reporting-key-insights-to-take-away-af17d985cdf2">PETs for Attribution and Reporting: Key Insights to Take Away</a> .</li></ul><p>In addition to the <a href="https://privacysandbox.com/">Privacy Sandbox</a> (PSB) by Google Chrome, <a href="https://github.com/patcg-individual-drafts/ipa/tree/main">Interoperable Private Attribution</a> (IPA) by Meta/Mozilla, and <a href="https://github.com/WICG/privacy-preserving-ads/tree/main">MaskedLARK</a> by Microsoft Edge, Apple recently published in September 2023 a novel proposal for reporting, coined <a href="https://github.com/patcg-individual-drafts/private-ad-measurement">Private Attribution Measurement</a> (PAM), which builds upon IPA and the former <a href="https://webkit.org/blog/11529/introducing-private-click-measurement-pcm/">Private Click Measurement</a> (PCM) proposal by Apple. PAM was first presented publicly at W3C TPAC in 2023, 11th September to the advertising industry.</p><p><strong>Objective.</strong> This article aims to introduce PAM, compare it with other browser vendors’ proposals, such as the Privacy Sandbox and IPA, deep-dive into its technical details, and propose tentative updates. In particular, we discuss potential future improvements that would match AdTech companies&#39; needs for both reporting and campaign optimization via machine learning training &amp; serving.</p><h3>TL;DR</h3><ul><li>From what it’s known about PAM, the design goal of this proposal by Apple is to support <em>at least</em> data queries that are already supported in the Attribution Reporting Aggregated (ARA) API of PSB. The main difference compared to ARA is that the <em>private</em> aggregation procedure is not performed via a trusted execution environment, but using secure 2-party computation (which does not suffer from side-channel attacks).</li><li>Similar to IPA, PAM allows to perform multiple aggregation queries on the same data (<em>e.g.,</em> impressions), involving potentially mutually exclusive breakdown keys. The privacy budget associated with differential privacy is split accordingly. This is on par with what Criteo envisions regarding the usage of aggregation APIs from browser vendors: AdTech companies should be able to explore and test several aggregation strategies on the same dataset for debugging and model learning purposes.</li><li>Again, similar to IPA and ARA (via the event-level reporting), PAM is supporting, in theory, both event-level and aggregated reporting, with a noise coming from differential privacy that might need to be adapted to the aggregation level to preserve user privacy. Nevertheless, we have some doubts about how the event-level reporting will work with the requirement for the AdTech company to list at conversion time all impression IDs associated with a particular conversion. The availability of event-level reporting is beneficial for AdTech companies willing to use such reports to train their machine learning models. Even if this type of reporting is not the objective of this proposal, we would be interested in pushing the idea of adapting the noise to the query; namely, would it be possible to have a large query choice and adapt the noise accordingly?</li></ul><h3>A high-level overview of Private Ad Measurement</h3><p>The content of this section is mainly based on the public information of PAM made available through a <a href="https://github.com/patcg-individual-drafts/private-ad-measurement">GitHub repository</a>.</p><h4>Overview</h4><p>PAM currently stands for a proposal that allows to measure, in a private manner, the effectiveness of advertisements made on the open internet. Here, privacy means user privacy and is defined as the inability of any advertising actor to track users across several websites. In its current form, PAM enables ad measurement through attribution and reporting. In other words, an advertiser using PAM would be able to measure where his advertising spend is the most effective by knowing the conversion value associated with a set of impressions. Its high-level design is depicted and detailed in Figures 1 and 2 below.</p><p><strong>Data storage &amp; attribution.</strong> Similar to ARA, attribution between an impression and a conversion event is performed on the user’s device. To do so, the browser is storing information about ad impressions in a database, as illustrated in Figure 1. Depending on their type, data are stored in two different databases:</p><ul><li><em>impression data</em> are stored at impression time in a database called “Advertiser“ in Figure 1. Such data refer to impression events and contain fields such as an <em>Ad ID key</em> set by the AdTech and referring to an aggregation breakdown key. This Ad ID key could, for instance, correspond to a campaign ID or a more granular identifier;</li><li><em>conversion data</em> are stored at conversion time into a different database, called “Publisher” in Figure 1. One row in this database corresponds to one conversion event, which involves a <em>Conversion ID key</em>. The AdTech company will provide PAM with a list of Ad ID keys that should be considered for each Conversion ID key along with an attribution logic (<em>e.g.</em> last-touch attribution), allowing to split the conversion value on each Ad ID key. The Conversion ID key could refer to a single or a group of users.</li></ul><blockquote><strong>An important note on the Ad ID key:</strong> From <a href="https://github.com/patcg-individual-drafts/private-ad-measurement/blob/main/private_ad_measurement_details.md#impression-data-recorded-on-device">PAM documentation</a>, the Ad ID key “could be a coarse identifier like a single ad campaign, one of two IDs to perform A/B comparisons, or more a granular identifier”. As far as we understand this Ad ID key, we believe that one of the main limitation of PAM compared to other proposals such as ARA is that the link between the Ad ID key and the Conversion ID should be set by the AdTech company at conversion time. This requirement seems doable and scalable for Ad ID keys such as campaign ID, but is however difficult to meet for an AD ID key standing for an impression ID, enabling event-level reporting. Indeed, for a given conversion on an advertiser website, the set of eligible impressions could be huge and this could impact the scalability of the reporting system proposed in PAM. Even if the goal of PAM is not supporting event-level reporting, we would be interested in knowing how further versions of PAM could handle easily event-level reporting.</blockquote><p>An example of a report for a single conversion event is depicted at the right hand of Figure 1 and corresponds to a histogram whose bins are Ad ID keys associated with a particular Conversion ID key. This computation happens on the device.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*W-DfGw4f8wNJOxYr5e0ihQ.png" /><figcaption>Figure 1 — On-device attribution using PAM. Source: <a href="https://github.com/patcg-individual-drafts/private-ad-measurement/blob/main/External%20Private%20Ad%20Measurement.pdf">https://github.com/patcg-individual-drafts/private-ad-measurement/blob/main/External%20Private%20Ad%20Measurement.pdf</a></figcaption></figure><blockquote><strong><em>Example:</em></strong> For the sake of clarity, we provide hereafter a concrete example of what could be a single conversion event produced by PAM. Let’s consider that an advertiser is launching a retargeting campaign, associated to an identifier <em>my_campaign_id</em>, and is willing to measure the performance of such a campaign via the number of sales that have been generated. At impression time, the advertiser will declare via PAM an Ad ID Key referring to <em>my_campaign_id</em>. At conversion time, the advertiser will define a Conversion ID which refers to all attributed sales that happened on her website. For a single impression, the conversion report produced by PAM will only be a single bin associated to my_campaign_ID and the conversion value will be the number of sales (integer).</blockquote><p><strong>Reporting. </strong>The reporting phase is illustrated in Figure 2 below. It mainly consists of aggregating all the previous conversion reports associated with the same Conversion ID key. Aggregation is, by default, a sum, but the final aggregated conversion value that is reported on each bin could refer to other statistics by modifying the conversion value (<em>e.g.,</em> by taking the square of it to report a statistics that could be used to compute the variance of the conversion value). This aggregation is performed privately via secure 2-party computation implemented using the <a href="https://crypto.stanford.edu/prio/">Prio framework</a>. As detailed in <a href="https://medium.com/criteo-engineering/an-introduction-to-pets-for-attribution-and-reporting-7e8cde13eca3">our previous article</a>, this private computation requires computing two secret shares of each conversion event, encrypting them, and sending them to each Prio node. The output of this computation is a histogram where the aggregated conversion value per bin is then obfuscated via global differential privacy (for instance, adding a Laplace noise to each value as in ARA). Still, this differential privacy noise addition is currently unspecified.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gxVl9O2HSyzC2hP8Py5QIA.png" /><figcaption>Figure 2 — Outsourced reporting using PAM. Source: <a href="https://github.com/patcg-individual-drafts/private-ad-measurement/blob/main/External%20Private%20Ad%20Measurement.pdf">https://github.com/patcg-individual-drafts/private-ad-measurement/blob/main/External%20Private%20Ad%20Measurement.pdf</a></figcaption></figure><h4>Main differences between IPA and ARA</h4><p>Table 1 below provides a high-level comparison between PAM, IPA, and ARA on two comparison points. More details can be found in <a href="https://docs.google.com/document/d/1oMe3Q7DMG3xzl-peIdq6voVD4PKD822-W4nf4ak7s2E/edit">this public document</a> presented at the W3C TPAC conference.</p><p>One of the main bottlenecks of PAM, compared to ARA, is the fact that PAM requires the device to send all encrypted histogram contributions even those that are equal to zero. As illustrated in the left-hand side of Figure 2, this is due to the fact that secure multi-party computation is working with secret shares standing for dense histograms even if the initial histogram is sparse.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3WBpXLitaKeSVzs0eh2PvQ.png" /><figcaption>Table 1 — High-level comparison between IPA, PAM and ARA.</figcaption></figure><h3>Thoughts on the future of reporting and learning using browser vendors&#39; proposals</h3><p>In this section, we propose to build upon the PAM proposal and its potential interoperability with other browser vendors’ proposals to highlight some important use-cases that such proposals could implement to meet the flexibility and performance requirements of AdTech companies, such as Criteo.</p><h4>Towards a common API surface with different backends</h4><p>Interestingly, PAM stands as a compromise between IPA and ARA in the sense that (1) PAM considers on-device attribution as ARA does; and (2) PAM builds upon secure multi-party computation to perform private computations on data collected by the user’s device, as IPA does. Since PAM and ARA are considering the same input data, namely encrypted histogram contributions of attributed conversions, this raises the question of whether these two proposals could become interoperable. Indeed, as highlighted in a recent W3C PATCG <a href="https://github.com/patcg-individual-drafts/private-ad-measurement/issues/6#issuecomment-1771679483">ad-hoc discussion</a>, it could be beneficial for the open internet to allow for different API backends depending on each browser&#39;s privacy requirements. As an example, we could imagine having a common API, similar to ARA or PAM, that would allow a browser (or even an AdTech company) to choose the backend behind the trusted server performing the private computation: either secure multi-party computation (MPC) <em>e.g.</em> via the Prio framework, or a trusted execution environment (TEE). For AdTech companies, such as Criteo, these different backends would be beneficial to arbitrate on the different reputational risks that a browser is willing to consider: collusion risk with MPC or side-channel attack risks with TEE as explained in <a href="https://medium.com/criteo-engineering/an-introduction-to-pets-for-attribution-and-reporting-7e8cde13eca3">our previous article</a>. In addition, the AdTech could also have the flexibility to choose more freely the TEE or MPC node providers depending on their partnerships or business considerations.</p><h4>On the need for multi-querying and privacy budget planning</h4><p>As Criteo already pointed out in <a href="https://github.com/WICG/attribution-reporting-api/issues/583">an ARA GitHub issue</a> for reporting use-cases, AdTech companies need to be allowed to span their differential privacy (DP) budget over time to make different queries on the same set of impressions. For example, for the campaign optimization use-case via machine learning training, an AdTech company might want to discover the utility of a set of ad features to optimize the performance of her bidding models. Imagine that the performance of a bidding model used to predict the probability of a sale given a landed click only depends on three features: the advertiser ID, the product ID, and the time since the last purchase of a user. If the AdTech company is spending all its privacy budget once on a different set of features, then no flexibility is given to the AdTech to correct this query in order to get more relevant features.</p><p>Up to our knowledge, this use-case could be related to <a href="https://github.com/patcg-individual-drafts/private-ad-measurement/issues/3">this PAM GitHub issue</a> but from a learning lens. We would be interested in knowing if such a use-case for campaign optimization is or could be supported by PAM in the near future. Some questions Criteo would be particularly interested in:</p><ul><li>Is there a DP budget (measured via the DP parameter epsilon) per impression as of today?</li><li>Would the AdTech be given the flexibility to split on her own the DP budget per impression so that the global DP guarantee is satisfied via query composition?</li></ul><h4>On the adaptivity of the noise with respect to the query</h4><p>Finally, we are building upon the fact that PAM is supporting, in theory, both aggregated and event-level reporting by allowing the Ad ID key to be as granular as possible. We would like to emphasize that such adaptivity of the noise with respect to the query is key for AdTech companies in order to find the queries that maximize the signal-to-noise ratio. We would be interested in knowing more about :</p><ul><li>the DP mechanisms that are used depending on the type of the conversion value (<em>e.g. </em>discrete versus continuous)?</li><li>an envisioned interoperability with the <a href="https://github.com/WICG/attribution-reporting-api/blob/main/flexible_event_config.md">flexible event-level reporting API of ARA</a> which would allow to adapt the noise to the support of the conversion value?</li></ul><p>Check all our articles about Privacy Sandbox and the future of “<em>attribution and reporting” </em>here 👇</p><p><a href="https://medium.com/criteo-engineering">Criteo R&amp;D Blog</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=52dcd41021f7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/criteo-engineering/apples-private-ad-measurement-52dcd41021f7">Apple’s Private Ad Measurement</a> was originally published in <a href="https://medium.com/criteo-engineering">Criteo Tech Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[PETs for Attribution and Reporting: Key Insights to Take Away]]></title>
            <link>https://medium.com/criteo-engineering/pets-for-attribution-and-reporting-key-insights-to-take-away-af17d985cdf2?source=rss-a90c05078adc------2</link>
            <guid isPermaLink="false">https://medium.com/p/af17d985cdf2</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[privacy-sandbox]]></category>
            <category><![CDATA[adtech]]></category>
            <category><![CDATA[advertising]]></category>
            <category><![CDATA[criteo]]></category>
            <dc:creator><![CDATA[Maxime Vono]]></dc:creator>
            <pubDate>Thu, 08 Jun 2023 06:37:15 GMT</pubDate>
            <atom:updated>2023-06-08T06:37:15.098Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cOFF8yqCE47RPD4MsDBwFA.jpeg" /></figure><h3>Introduction</h3><p>As a leading AdTech company that drives commerce outcomes for media owners and marketers, Criteo is one of the companies shaping the industry’s future, notably on attribution and reporting. As such, we published a series of articles focusing on understanding and discussing the technologies underlying current browser vendors’ proposals for <em>attribution and reporting</em> use cases:</p><ul><li><a href="https://medium.com/criteo-engineering/an-introduction-to-pets-for-attribution-and-reporting-7e8cde13eca3">An Introduction to PETs for Attribution and Reporting</a> ;</li><li><a href="https://medium.com/criteo-engineering/pets-in-advertising-scenarios-for-secure-multi-party-computation-4d2e667e5ecd">PETs in Advertising: Scenarios for Secure Multi-Party Computation</a> ;</li><li><a href="https://medium.com/criteo-engineering/pets-in-advertising-scenarios-for-trusted-execution-environments-9d0264c57325">PETs in Advertising: Scenarios for Trusted Execution Environments</a> .</li></ul><p>Such technologies are referred to as <em>privacy-enhancing technologies</em> (PETs) and are being proposed to preserve user privacy while allowing cross-site measurement and attribution. To meet this goal, two PETs are currently being investigated by major browser vendors (<a href="https://privacysandbox.com/">Google Chrome</a> &amp; <a href="https://github.com/patcg-individual-drafts/ipa/blob/main/IPA-End-to-End.md">Meta/Mozilla Firefox</a>), namely <em>secure multi-party computation </em>(MPC) and <em>trusted execution environments</em> (TEEs).</p><p>This article summarises our main findings regarding the cost of using such PETs for standard advertising use cases, requirements for AdTech companies willing to use APIs leveraging these PETs, and opportunities AdTech companies might unlock beyond current proposals.</p><h3>TL;DR</h3><ul><li>For simple advertising use cases (<em>e.g.</em> attribution &amp; reporting), leveraging either MPC or TEE does not yield significant costs for AdTech companies: the relative cost compared to the campaign cost is of order 0.6% in the least favorable scenario.</li><li>More sophisticated use cases (e.g. campaign optimisation via machine learning training) are currently not supported in browser vendors’ proposals. Performing these use cases directly within TEE-based trusted servers is doable but further investigation is needed to answer how the audit of the code within the TEE will be done. We hope that this series of articles will allow us to foster discussions in the advertising landscape to address the aforementioned issues.</li><li>Finally, we identified areas of improvement regarding browser vendors’ proposals to empower flexibility, performance, and cost efficiency of AdTech companies. Notably, we would like to initiate discussions (<em>e.g. </em>see our <a href="https://github.com/patcg/meetings/issues/122">github issue</a>) on the possibility for AdTech companies to host TEEs on-premise to run bidding and reporting Privacy Sandbox APIs.</li></ul><h3>PET-based browser vendors’ proposals for Attribution and Reporting</h3><p>As summarised in the below table, several browser vendors have made proposals that leverage PETs to perform attribution and reporting. Note that, given the large number of browser vendors’ proposals (including those not leveraging MPC or TEE), we are not committed to performing an comprehensive analysis of all these proposals. Nevertheless, some of our findings might be re-used to understand complementary browser vendors’ proposals such as PARAKEET from Microsoft Edge [6–7].</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IrFBcJllRvkH8CebStXysg.png" /><figcaption><em>Table 1 — Summary of browser vendors’ proposals / APIs based on PETs.</em></figcaption></figure><h3>Cost of PET-powered advertising use cases</h3><p>We gather in Table 2 below envisioned computing and networking costs for AdTech companies when calling PET-based APIs to perform attribution, reporting, and campaign optimisation. Note that other types of cost should also be considered such as, when considering a TEE, the cost of attesting the code running within the TEE, or the hardware expenditure cost if the TEE is running on-premise.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tnj5EYpITOg3-qwDpg0RMA.png" /><figcaption><em>Table 2 — Envisioned costs associated to PET-enhanced advertising use-cases.</em></figcaption></figure><h3>Considerations for AdTech companies in the status quo</h3><p>This section summarises some key considerations that AdTech companies should factor in when implementing current proposals.</p><p><strong>Trust</strong></p><ul><li>For MPC-based browser vendors’ proposals, AdTech companies have to trust that a set of <em>trusted helper servers</em> owned by third-party companies will not collude <em>i.e. </em>will not share their private input information among themselves. Otherwise, these helper servers will be able to perform cross-domain tracking of individual users.</li><li>For TEE-based browser vendors’ proposals, AdTech companies will need to trust both the TEE manufacturer (<em>e.g.</em> Intel or AMD) — having no prior interest in advertising nor marketing — and the party hosting the TEE (<em>e.g.</em> the cloud vendor).</li></ul><p><strong>Trusted server hosting</strong></p><p>Notably for TEE-based proposals, the AdTech company will be required to run TEEs in cloud instances. Being required to operate on a cloud platform is likely to increase an AdTech company’s costs significantly, from those operating from their own infrastructure such as Criteo. As far as we know, a majority of the largest Supply and Demand Side Platforms are running part if not all on premise, or are actually cloud providers [8–9]. It is therefore critical that browser vendors express the infrastructure and security requirements under which AdTech companies could perform reporting and bidding on-premise, rather than defaulting to cloud as a de facto hosting solution. To initiate these discussions, we opened <a href="https://github.com/patcg/meetings/issues/122">an agenda request</a> to have the possibility to discuss on-premise TEE hosting in the upcoming W3C event in London.</p><p><strong>Machine learning training via trusted helper servers</strong></p><p>Currently, campaign optimisation via the usage of trusted helper servers is not supported by major browser vendors’ proposals such as Interoperable Private Attribution of Meta/Mozilla and the Privacy Sandbox of Google Chrome. While event-level reports might be available for some years via Privacy Sandbox reporting APIs, resorting to PET-enhanced trusted servers leaves open the opportunity for AdTech to perform ML training directly within these trusted servers. In particular, this possibility would allow AdTech companies to:</p><ul><li>Innovate by training sophisticated ML algorithms on granular data (<em>e.g.</em> deep learning approaches) instead of devising specific and challenging ML approaches for aggregated input data;</li><li>Reduce the quantity of noise added by resorting to global differential privacy instead of local differential privacy.</li></ul><p>Such opportunity is also pushed by other actors in the advertising landscape (<em>e.g.</em> see <a href="https://github.com/patcg/meetings/issues/117">this github issue</a> or MaskedLARK proposal by Microsoft Edge). We believe that a standardisation of how to solve the “private campaign optimisation” use case is important and we are looking forward to participate in upcoming discussions on this topic.</p><h3>What’s Next</h3><p>We would like to present some of our findings in an upcoming W3C event in London on 14–15 June, 2023. We also plan to publish another article on how trusted execution environments could be leveraged to address a variety of issues regarding bidding and auctions, such as avoiding auction replay.</p><h3>References</h3><p>[Criteo 2] <a href="https://medium.com/criteo-engineering/pets-in-advertising-scenarios-for-secure-multi-party-computation-4d2e667e5ecd">PETs in Advertising: Scenarios for Secure Multi-Party Computation</a></p><p>[Criteo 3] <a href="https://medium.com/criteo-engineering/pets-in-advertising-scenarios-for-trusted-execution-environments-9d0264c57325">PETs in Advertising: Scenarios for Trusted Execution Environments</a></p><p>[1] <a href="https://privacysandbox.com/">Privacy Sandbox</a></p><p>[2] <a href="https://github.com/WICG/attribution-reporting-api/blob/main/AGGREGATION_SERVICE_TEE.md">Aggregated Attribution Reporting API</a></p><p>[3] <a href="https://github.com/privacysandbox/fledge-docs/blob/main/bidding_auction_services_api.md">Bidding and Auctions Services API</a></p><p>[4] Protected Audience API’s <a href="https://github.com/privacysandbox/fledge-docs/blob/main/key_value_service_trust_model.md">key/value server</a></p><p>[5] <a href="https://github.com/private-attribution/ipa">Interoperable Private Attribution</a></p><p>[6] <a href="https://github.com/WICG/privacy-preserving-ads/blob/main/Parakeet.md">PARAKEET</a></p><p>[7] <a href="https://github.com/WICG/privacy-preserving-ads/blob/main/MaskedLARK.md">MaskedLARK</a></p><p>[8] <a href="https://pubmatic.com/blog/building-for-the-future-of-adtech-with-a-private-cloud-data-center-strategy-tech-blog/">Building a Future with Private Cloud Computing for Advertising Technology</a></p><p>[9] <a href="https://investor.magnite.com/news-releases/news-release-details/magnite-invests-new-data-center-singapore-support-growing-apac">Magnite Invests In New Data Center in Singapore to Support Growing APAC Business | Magnite, Inc.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=af17d985cdf2" width="1" height="1" alt=""><hr><p><a href="https://medium.com/criteo-engineering/pets-for-attribution-and-reporting-key-insights-to-take-away-af17d985cdf2">PETs for Attribution and Reporting: Key Insights to Take Away</a> was originally published in <a href="https://medium.com/criteo-engineering">Criteo Tech Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[PETs in Advertising: Scenarios for Secure Multi-Party Computation]]></title>
            <link>https://medium.com/criteo-engineering/pets-in-advertising-scenarios-for-secure-multi-party-computation-4d2e667e5ecd?source=rss-a90c05078adc------2</link>
            <guid isPermaLink="false">https://medium.com/p/4d2e667e5ecd</guid>
            <category><![CDATA[criteo]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[privacy]]></category>
            <category><![CDATA[privacy-sandbox]]></category>
            <category><![CDATA[adtech]]></category>
            <dc:creator><![CDATA[Maxime Vono]]></dc:creator>
            <pubDate>Thu, 04 May 2023 06:21:48 GMT</pubDate>
            <atom:updated>2023-05-04T06:49:05.019Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sfBQRpjDFgeFVxyxLuMhvw.jpeg" /></figure><h3>Introduction</h3><p><strong>Context.</strong> In our <a href="https://medium.com/criteo-engineering/an-introduction-to-pets-for-attribution-and-reporting-7e8cde13eca3">previous article</a>, we introduced and compared in a high-level manner two privacy-enhancing technologies (PETs) that are currently being investigated [1-2] by major browser vendors (<em>e.g.</em> Google Chrome &amp; Mozilla Firefox) to address reporting use cases while meeting <a href="https://github.com/patcg/docs-and-reports/tree/main/threat-model">privacy guarantees</a>. Among these PETs, secure multi-party computation (MPC) [3] is currently explored in the Interoperable Private Attribution(IPA) [1] proposal to perform private attribution and reporting.</p><p><strong>Objective.</strong> This article aims to deep-dive into the technical details of MPC for advertising use cases including private attribution, reporting and campaign optimisation which have been introduced in our <a href="https://medium.com/criteo-engineering/an-introduction-to-pets-for-attribution-and-reporting-7e8cde13eca3">previous article</a>. More precisely, we will give a high-level overview of IPA and detail the main building blocks that would be leveraged to address each use case using MPC. In addition, we will provide you with insights into the complexity, scalability and expected cost associated with each use case if performed via MPC. The latter will be finally illustrated via realistic numerical experiments on relevant advertising data sets, including open-source ones.</p><h3>TL;DR</h3><ul><li>From the recent performance benchmarks of IPA [4,5], the relative costs of performing attribution and reporting using MPC compared to the campaign cost are roughly 0.6% and 0.2%, respectively.</li><li>From our own benchmarks, campaign optimisation via machine learning (ML) training using MPC is 180 times more costly (in terms of computing cost) than ML training without MPC, if the same number of training samples is used. This overhead might be reduced using a better implementation of MPC protocols but this remains an open question for the moment.</li><li>Compared to trusted execution environments discussed in the <a href="https://medium.com/criteo-engineering/an-introduction-to-pets-for-attribution-and-reporting-7e8cde13eca3">previous article</a>, MPC provides stronger privacy guarantees in theory since no side-channel attacks could be leveraged. Nevertheless, MPC could suffer from other threats such as sharing agreements across helper parties as pointed out in our <a href="https://medium.com/criteo-engineering/an-introduction-to-pets-for-attribution-and-reporting-7e8cde13eca3">previous article</a>.</li></ul><h3>A high-level overview of IPA</h3><h4>Overview</h4><p>Before deep-diving into the technical details of how MPC could be used to perform private attribution, reporting and campaign optimisation, let’s start with a high-level picture of the IPA proposal along with some potential additional components (e.g. campaign optimisation).</p><p>Figure 1 below depicts an example of a data processing pipeline that could be performed using IPA. This pipeline is detailed in the sequel.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4StK2RbnajUVpEUN7EtTWA.png" /><figcaption>Figure 1 — From fragmented source (e.g. publisher) and trigger (e.g. advertiser) events to reporting and campaign optimisation, passing through attribution.</figcaption></figure><p>We consider that a user is browsing sequentially two types of websites (or mobile applications). First, the user browses a so-called <em>source </em>website (e.g. a publisher website) on which s/he interacts leading to <em>source events</em> (e.g. a click on an ad banner). Then, the user lands on a <em>trigger </em>website (e.g. an advertiser website) on which s/he decides to buy a product, for instance, leading to a sale.</p><p>An AdTech company (called <em>report collector</em> in IPA), willing to measure and optimise advertising campaigns, is assumed to be able to collect two types of reports: the <em>source report</em> and the <em>trigger report</em>. In order to avoid cross-domain tracking, the ID of the user (referred to as the <em>match key</em> in the IPA proposal) is encrypted and the encryption function will be different for each domain. The encryption process proposed in IPA is based on secret sharing and asymmetric encryption and will be detailed in the next paragraph. Attribution, via ID resolution, will be performed by a helper party network involving at least two helper parties. As highlighted in Figure 1, reporting and campaign optimisation use cases require attribution of trigger events to source events. For reporting, this first step is necessary to be able to aggregate trigger events per source dimension (e.g. a positive click as shown in Figure 1). For campaign optimisation, this step allows you to build a training data set involving a set of features (e.g. a click) and labels (e.g. a sale) needed to train a machine learning (ML) model.</p><h4>Helper party network and secret sharing</h4><p>As introduced in our <a href="https://medium.com/criteo-engineering/an-introduction-to-pets-for-attribution-and-reporting-7e8cde13eca3">previous article</a> (see <em>Secure Multi-Party Computation</em> paragraph), IPA and more generally MPC are based on two key ingredients to meet privacy requirements: a <strong>helper party network</strong> made of several helper parties and <strong>secret sharing</strong>. As in IPA, we will assume in this article that three helper parties are available.</p><p>Secret sharing is performed by two actors, at different moments and on different data.</p><ol><li>Assuming that an ID (a.k.a. match key) has been provided to the browser, the latter creates three shares of the ID, encrypts them using public keys provided by helper parties (e.g. share 1 encrypted using a public key of helper party 1), and sends the encrypted shares to the AdTech company. Along with these encrypted shares, the AdTech company also receives source and trigger plaintext data of events, as shown in Figure 1.</li><li>To prevent helper parties from learning new information, the AdTech company creates three shares of each plaintext data attribute and encrypts them with helper parties’ public keys. Then, the AdTech company sends both encrypted IDs and attribute shares to helper parties.</li></ol><p>To sum up, after decrypting their encrypted shares, each helper party owns a secret share of all the input data (IDs and attributes). MPC protocols to perform private attribution, reporting and campaign optimisation can now be applied. They are detailed in the following sections.</p><p>We would like to emphasise that we do not aim to present all surrogate approaches to cope with threat models detailed in the PATCG document [1] and all IPA protocols presented in the companion research paper [6]. Instead, we are focusing on presenting the main details of MPC protocols associated with attribution and reporting.</p><h3>Private attribution via MPC</h3><p>Standard attribution between source and trigger events stands for an INNER JOIN operation on a common match key. This JOIN operation is typically performed via hash or sort-merge join algorithms [7,8]. The latter class of approaches is usually preferred when memory to store the hash table is limited; and consists in first sorting the table per match key before scanning it. Compared to traditional sort-merge join, private attribution using MPC involves an additional constraint to avoid input leakage: the sequence of comparisons to sort rows should not depend on the input. This leads to considering an enhanced class of sorting algorithms referred to as <strong>oblivious sorting [</strong>9]. Since the sorting operation is the main bottleneck of join procedures, we only focus on this part in this section.</p><p>Many oblivious sorting approaches have been proposed in the literature such as <a href="https://arxiv.org/abs/0909.1037">Shellsort</a>, <a href="https://arxiv.org/abs/1403.2777">Zigzag sort</a>, <a href="https://link.springer.com/chapter/10.1007/978-3-319-11599-3_4">radix sort in the context of the Sharemind system</a>, <a href="https://eprint.iacr.org/2022/1595.pdf">enhanced versions of secure radix sort</a> or <a href="https://link.springer.com/chapter/10.1007/978-3-642-37682-5_15">oblivious Quicksort</a>. These works mainly focus on improving the constant in the asymptotic lower bound O(<em>m</em>.log(<em>m</em>)) regarding communication costs, where <em>m</em> stands for the total number of rows to be sorted. Table 1 below depicts the communication overhead of a state-of-the-art oblivious sorting protocol, currently considered in IPA, under both semi-honest and malicious security.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zBhpdJoNNwz-G8knF6NBeg.png" /><figcaption>Table 1. Comparison between standard sorting and secure sorting via MPC [10]. The variable <em>l</em> stands for the bit length of the match key, <em>m</em> for the total number of rows to be sorted, <em>l</em>’ for the bit length of other columns (e.g. trigger event) and <em>e</em> for the cheating/failure probability in a malicious scenario. In the last column, we only compared the main term in O(<em>l</em>.<em>m</em>.log(<em>m</em>)), and used <em>m</em> = 1,000,000 and<em> e</em> = 2–7 (corresponding to a statistical security parameter of 8 bits).</figcaption></figure><p>Once the secret-shared table has been sorted, one can perform last-touch attribution via a tree-like procedure. This can be achieved with O(log(<em>m</em>)) communication cost, significantly lower than the overhead associated with the sorting phase.</p><p>Based on IPA performance benchmarks performed by Meta/Mozilla [13,14], Table 2 below illustrates the costs associated with the sorting and attribution stages for two implementations of IPA: based on an MPC engine called MP-SPDZ [15] and a production-oriented implementation on Rust. The associated total cost is then compared to an average $2 CPM which could be a reasonable assumption considering worldwide advertising channels. Obviously, these figures could vary depending on cloud providers’ costs, the location of the cloud instances and the true auction cost to show ad impressions. One can denote that significant improvement regarding the network cost has been obtained via a Rust implementation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HhsSbKeP6sUWIh_y8ayyQA.png" /><figcaption>Table 2. Hypothetical cost of the IPA Reporting API by considering 3 m5dn.12xlarge instances that cost $3.2/hour each, an average CPM of $2, and a network cost of $0.08/GB [4,5]. These costs have been measured using 1,000,000 rows, 16 breakdown keys and 40-bit match keys under the malicious adversarial model. “Estimate” means that total cost has been estimated using <em>computing cost</em> = 0.5 * <em>network cost</em>.</figcaption></figure><h3>Private reporting via MPC</h3><p>Once the secret-shared trigger and source events have been attributed, one can perform aggregation by, for instance, summing all trigger values per source dimension (referred to as <em>breakdown key</em> in IPA). This step could be performed similarly to the previous sorting and attribution procedure by first sorting the secret-shared attributed events by breakdown keys and then using a prefix-sum algorithm to aggregate the trigger values per breakdown key. Since the number of breakdown keys would be significantly lesser than the number of rows, we could expect that the communication overhead associated with this step is less than that of the oblivious sorting step, which is confirmed experimentally by performance benchmarks performed by IPA team [4,5], see also Figure 2 below. Note that these results are associated with a research prototype that might be further improved. Nonetheless, they still provide an idea of the costs associated with the use of MPC for attribution and reporting.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LKXnExYo6E2pBf4YfRV6VQ.png" /><figcaption>Figure 2. Cost repartition for the IPA Attribution API using MP-SPDZ (August 2022) and Rust (February 2023) implementations.</figcaption></figure><p>As in Table 2 above, Table 3 depicts the cost of Reporting stage in IPA Reporting API that could be envisioned under some assumptions regarding the network and computing costs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vdzp6H1HF9kNBmYl3eJoLw.png" /><figcaption>Table 2. Hypothetical cost of the IPA Reporting IPA by considering 3 m5dn.12xlarge instances that cost $3.2/hour each, an average CPM of $2, and a network cost of $0.08/GB.</figcaption></figure><h3>Private campaign optimisation via MPC</h3><p>Finally, we would like to go beyond the IPA proposal and highlight <strong>how</strong>, and <strong>at which cost</strong>, MPC could be leveraged to perform private campaign optimisation via the training of a machine learning model by the helper party network.</p><p>To this end, we will focus on training a simple supervised classification model <em>using a Criteo open-source relatively small dataset </em>to estimate the probability of a click on an ad banner given a display [11-13].</p><p><strong>Dataset.</strong> We used an open-source Criteo dataset publicly available at <a href="http://go.criteo.net/criteo-ppml-challenge-adkdd21-dataset-raw-granular-data.csv.gz">http://go.criteo.net/criteo-ppml-challenge-adkdd21-dataset-raw-granular-data.csv.gz</a>. It has been used for our <a href="https://competitions.codalab.org/competitions/31485#learn_the_details-technical-details">KDD privacy challenge</a>. This dataset has ~89M observations and 19 categorical features with many modalities. The target binary variable is the click column.</p><p><strong>Pre-processing.</strong> To reduce overhead, feature and observation sampling have been performed. More precisely, the original 12th and 14th columns have been removed and only the most frequent 5,000 features after one-hot encoding and hashing have been kept. In addition, only 1M observations have been used for training (in production, note that we collect 1 billion observations per day). Despite feature reduction, the selected 5,000 features are still sparse and this sparsity cannot be efficiently coped with by MPC to perform matrix-vector product computations since this would leak the position of the zeros. To address this problem, a linear projection of the 5,000 features has been performed leading to dense embeddings of size 4.</p><p><strong>Model. </strong>For the classification model, we have chosen a small shallow neural network with one hidden layer of size 128 and ReLU activation function. The sigmoid function stands for the activation function of the output layer.</p><p><strong>Loss. </strong>The training loss stands for the binary cross-entropy also referred to as log-loss.</p><p><strong>MPC protocol. </strong>This model has been trained using three helper parties, each running an Amazon AWS c6i.2xlarge instance in the us-west-1 time zone. MPC protocols from <a href="https://eprint.iacr.org/2018/403.pdf">ABY3</a> have been used under the semi-honest security model. We did not perform experiments under malicious security for the sake of simplicity and since the semi-honest security model already provides a lower bound on the computation cost. Non-linear functions in the training loss function, namely the logarithm, sigmoid and ReLU functions, have been approximated. As an example, for the sigmoid function, one could leverage approximations based on piecewise polynomials as detailed in <a href="https://eprint.iacr.org/2018/403.pdf">ABY3</a> or exponential and reciprocal function approximations. Numerical illustrations of these approximations can be found in <a href="https://github.com/LaRiffle/approximate-models/blob/master/experiments/Logarithm.ipynb">this notebook</a> and mathematical details are provided in this <a href="https://arxiv.org/pdf/2109.00984.pdf">paper</a> and in <a href="https://eprint.iacr.org/2018/403.pdf">ABY3</a>. Implementation has been performed using the <a href="https://github.com/ciphermodelabs/ciphercore">CipherCore MPC engine</a>, mostly coded in Rust.</p><p><strong>Training details. </strong>We used<strong> </strong>stochastic gradient descent with a batch size of 512, Adagrad optimizer with a learning rate 0.05, fixed-precision arithmetic with 15 bits after the decimal point.</p><p><strong>Results.</strong> Those are depicted in Table 3 below. For the sake of comparison, a baseline logistic regression model, using 89M observations and 216 features (to be compared to roughly 1B features in our production models) after one-hot encoding and hashing, has been trained via stochastic gradient descent and Adam optimizer on an Amazon AWS c6i.2xlarge instance in the us-west-1 time zone.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MeBgrnVtlEIfzbFZrbeq0Q.png" /><figcaption>Table 3. Comparison between standard logistic regression (LR) training and the proposed shallow neural network (NN) training with MPC. AUC stands for <em>area under curve</em>. <br>*The naive model always outputs the mean label value of the dataset, in our case 0.1.</figcaption></figure><p>Although training this shallow neural network on a reduced set of features and training observations yields acceptable performance at a reasonable computing cost, one should bear in mind that the number of features and size of the dataset (1M observations) is very small compared to standard advertising training logs (1 billion observations). In addition, recall that all experiments have been run under the semi-honest security model so we expect additional overhead under malicious security. It would be interesting to see if we could obtain reasonable network and computing costs using a more efficient implementation of MPC protocols for ML training. For the moment, this question remains open.</p><h3>What’s next</h3><p>In our next article, Criteo will present you in detail how attribution, reporting and campaign optimisation could be performed using TEE-based trusted servers, and at which envisioned cost.</p><h3>References</h3><p>[1] <a href="https://github.com/patcg-individual-drafts/ipa/blob/main/IPA-End-to-End.md">ipa/IPA-End-to-End.md at main · patcg-individual-drafts/ipa</a></p><p>[2] <a href="https://github.com/WICG/attribution-reporting-api/blob/main/AGGREGATE.md">attribution-reporting-api/AGGREGATE.md at main · WICG/attribution-reporting-api</a></p><p>[3] <a href="https://en.wikipedia.org/wiki/Secure_multi-party_computation">Secure multi-party computation</a>; <a href="https://eprint.iacr.org/2020/300.pdf">https://eprint.iacr.org/2020/300.pdf</a></p><p>[4] <a href="https://github.com/patcg/meetings/blob/main/2023/02/06-nyc/PATCG%20Feb%202023%20-%20IPA%20Update.pdf">meetings/PATCG Feb 2023 — IPA Update.pdf at main · patcg/meetings</a></p><p>[5] <a href="https://github.com/patcg/meetings/blob/main/2022/08/09-telecon/IPA-August-2022-Update-PATCG-Issue-70.pdf">meetings/IPA-August-2022-Update-PATCG-Issue-70.pdf at main · patcg/meetings</a></p><p>[<a href="https://eprint.iacr.org/2023/437.pdf">6] https://eprint.iacr.org/2023/437.pdf</a></p><p>[7] <a href="https://dl.acm.org/doi/10.14778/1687553.1687564">https://dl.acm.org/doi/10.14778/1687553.1687564</a></p><p>[8] <a href="https://ieeexplore.ieee.org/abstract/document/5388057">https://ieeexplore.ieee.org/abstract/document/5388057</a></p><p>[9] <a href="https://eprint.iacr.org/2011/122.pdf">https://eprint.iacr.org/2011/122.pdf</a></p><p>[10] <a href="https://eprint.iacr.org/2022/1595.pdf">https://eprint.iacr.org/2022/1595.pdf</a></p><p>[11] <a href="https://people.csail.mit.edu/romer/papers/TISTRespPredAds.pdf">https://people.csail.mit.edu/romer/papers/TISTRespPredAds.pdf</a></p><p>[12] <a href="http://quinonero.net/Publications/predicting-clicks-facebook.pdf">http://quinonero.net/Publications/predicting-clicks-facebook.pdf</a></p><p>[13] <a href="https://jmlr.org/papers/volume15/agarwal14a/agarwal14a.pdf">https://jmlr.org/papers/volume15/agarwal14a/agarwal14a.pdf</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4d2e667e5ecd" width="1" height="1" alt=""><hr><p><a href="https://medium.com/criteo-engineering/pets-in-advertising-scenarios-for-secure-multi-party-computation-4d2e667e5ecd">PETs in Advertising: Scenarios for Secure Multi-Party Computation</a> was originally published in <a href="https://medium.com/criteo-engineering">Criteo Tech Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[An Introduction to PETs for Attribution and Reporting]]></title>
            <link>https://medium.com/criteo-engineering/an-introduction-to-pets-for-attribution-and-reporting-7e8cde13eca3?source=rss-a90c05078adc------2</link>
            <guid isPermaLink="false">https://medium.com/p/7e8cde13eca3</guid>
            <category><![CDATA[privacy-sandbox]]></category>
            <category><![CDATA[advertising]]></category>
            <category><![CDATA[adtech]]></category>
            <category><![CDATA[privacy]]></category>
            <category><![CDATA[criteo]]></category>
            <dc:creator><![CDATA[Maxime Vono]]></dc:creator>
            <pubDate>Tue, 25 Apr 2023 06:47:31 GMT</pubDate>
            <atom:updated>2023-04-25T06:47:31.496Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cLiG91-uwsMllAkUJYKXiQ.jpeg" /></figure><h3>Introduction</h3><p><strong>Context.</strong> To avoid direct cross-site tracking, several browsers (<em>e.g. </em><a href="https://github.com/patcg-individual-drafts/ipa/blob/main/IPA-End-to-End.md">Mozilla</a>, <a href="https://github.com/WICG/attribution-reporting-api/blob/main/AGGREGATE.md#data-processing-through-a-secure-aggregation-service">Google Chrome</a>, and <a href="https://github.com/WICG/privacy-preserving-ads/blob/main/NoisyRanking.md">Microsoft</a>) are developing <em>attribution</em> and <em>reporting</em> proposals that leverage the concept of <em>private computation</em> under the coordination of trusted helper servers. The latter concept aims at performing attribution and evaluating a pre-defined function (e.g. counting trigger events such as clicks) without revealing any new information to each party except the output of the aforementioned function.</p><p>Currently, two types of private computation systems based on <em>privacy-enhancing technologies </em>(PETs) are considered in browsers’ proposals. The first one, suggested in the <a href="https://github.com/patcg-individual-drafts/ipa/blob/main/IPA-End-to-End.md">Interoperable Private Attribution</a> (IPA) by Meta/Mozilla, is called <em>secure multi-party computation</em> (MPC) [1]. The second one, proposed in the <a href="https://github.com/WICG/attribution-reporting-api/blob/main/AGGREGATE.md#data-processing-through-a-secure-aggregation-service">Aggregate Reporting API</a> by Google, is based on <em>trusted execution environments </em>(TEEs) [2].</p><p><strong>Objective.</strong> In this article, we will guide you through these two PETs. Our objective is to provide a high-level understanding of the technologies underlying current proposals for <em>attribution and reporting</em> use cases, discuss their benefits but also their limitations, and foster cross-collaboration to address the areas of improvement.</p><p><strong>Related Works.</strong> As a leading AdTech company that drives commerce outcomes for media owners and marketers, Criteo is committed to evaluating proposals that might affect the way we will perform <em>attribution</em> and <em>reporting</em> in the future. Criteo has already participated in testing and providing feedback on browser proposals such as the <a href="https://privacysandbox.com/">Privacy Sandbox</a> one; see our previous two articles on <a href="https://developer.chrome.com/docs/privacy-sandbox/topics/">Topics API</a> and <a href="https://developer.chrome.com/docs/privacy-sandbox/fledge/">FLEDGE</a>:</p><ul><li><a href="https://medium.com/criteo-engineering/is-googles-topics-api-a-viable-replacement-for-interest-based-advertising-297076192bd">Topics API: Criteo’s First Look at Google’s Interest-Based Advertising Solution</a>;</li><li><a href="https://medium.com/criteo-engineering/an-update-on-fledge-chrome-testing-d0046430a3ec">An Update on FLEDGE Chrome Testing</a>.</li></ul><p>This article stands as the first introductory blog post among a series of articles focused on evaluating MPC and TEEs, notably for <em>attribution</em> and <em>reporting</em>.</p><h3>TL;DR</h3><ul><li>The usage of MPC or a TEE to perform private computation leads to considering two different trust paradigms. Regarding MPC, which involves multiple parties, we have to trust the latter to not share their private inputs among themselves. In contrast, building upon a hardware-based TEE avoids the necessity to perform multi-party computation; but trust is delegated to the hardware manufacturer in most cases.</li><li>A TEE involves a <em>remote attestation</em> feature that enables any party to audit the code (<em>e.g.</em> aggregation procedure for <em>reporting</em>) running inside it before sending data, contributing to ensuring the <em>correctness</em> of the private computation. This feature can be made available for MPC protocols if the latter is run inside a TEE.</li><li>Regarding the <em>accuracy</em> of the private computation, TEEs seem to stand for a better alternative compared to MPC. The <em>scalability</em> of such PET-based private computation remains detrimental for AdTech; it is for the moment not clear if TEEs and MPC yield reasonable computational and communication costs for standard use cases. This will be the focus of companion articles within this series.</li></ul><h3>High-level Design</h3><p>Browser vendors’ proposals such as the <a href="https://github.com/patcg-individual-drafts/ipa/blob/main/IPA-End-to-End.md#parties-involved">Interoperable Private Attribution</a> (IPA) by Meta/Mozilla or the <a href="https://github.com/WICG/attribution-reporting-api/blob/main/AGGREGATE.md">Attribution Reporting API</a> by Google Chrome, partially rely on so-called <em>helper parties </em>to instantiate a private computation procedure.</p><p>Delegating computation to these helper parties assumes that we can trust them in guaranteeing at least the following two properties, as proposed by the <a href="https://patcg.github.io/">PATWG</a> of the <a href="https://www.w3.org/">W3C</a> [3]:</p><ol><li><strong>Private computation:</strong> Helper parties cannot access raw user-level data.</li><li><strong>Correctness:</strong> Helper parties execute the prescribed protocol correctly.</li></ol><p>In Figure 1 below, we illustrate the main differences between helper parties leveraging a TEE and those instantiating an MPC protocol. <br>With TEEs, only one helper party is involved. Encrypted reports gathered by AdTech are sent to the TEE, which decrypts them in a secure environment and performs <em>attribution</em> and <em>reporting</em> on plaintext data. <br>With MPC protocols, several helper parties are involved. Encrypted reports are split into several random pieces called <em>shares </em>(see paragraph on <em>Secure Multi-Party Computation</em>) and then sent to several helper parties which decrypt them and perform computation on random pieces of data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NI6Csqqi_usvJXu0MDRxgw.png" /><figcaption>Figure 1. Private Computation using PET-based helper parties.</figcaption></figure><p>In the following paragraphs, we introduce MPC and TEEs and explain how the latter is proposed to be leveraged to meet <em>the private computation </em>and <em>correctness </em>properties.</p><h3>Trusted Execution Environment</h3><p>A <em>trusted execution environment</em> (TEE) provides trust and confidentiality guarantees for data in use; similar to transport layer security for data in transit or encryption algorithms for data at rest. A TEE stands for an isolated processing environment, based on specific hardware, in which applications can be securely executed in isolation of the operating system or hypervisor as illustrated in Figure 2.</p><p>Thanks to the isolation property provided by the TEE, the code can be processed on plaintext data inside the TEE.</p><p>Loosely speaking, TEEs can be divided into two categories depending on what the TEE executes:</p><ul><li>Process-based TEEs that only run a specific piece of code associated with an application in a secure manner;</li><li>VM-based TEEs that run the whole code of an application.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*S-4OweF-4aXlXaN2Y-3lNQ.png" /><figcaption>Figure 2. Illustration of the two main categories of TEEs (last two columns). The first column stands for a standard processor. “Trusted, but verified” means that trust is ensured via a verification process, called attestation.</figcaption></figure><p>Let’s now go back to the two PATCG properties: <em>private computation</em> and <em>correctness</em>.</p><p><strong>Private computation</strong> is achieved by guaranteeing that parties running computations in a TEE cannot tamper with nor access the code and data loaded by the TEE. In practice, this property is partially ensured through two means:</p><ul><li>reducing the communication interface, referred to as the trusted computing base of the TEE with the rest of the system so that the attack surface is significantly reduced;</li><li>encrypting values in memory.</li></ul><p>Despite these properties, potential leakage might still be possible by performing side-channel attacks via, for instance, exploitation of access patterns. This will be further detailed in an upcoming article focusing on TEEs.</p><p><strong>Correctness.</strong> On the other hand, <em>correctness</em> is achieved via a <em>remote attestation</em> feature which ensures (1) that the code being executed runs inside a genuine TEE instance <em>i.e.</em> created via specific hardware with confidential computing capabilities such as Intel SGX or AMD SEV; and (2) that the code is the expected one. Remote attestation is generally achieved by having the code digitally signed with a manufacturer-derived private key. Any party can then authenticate the code and verify it was not tampered with using the corresponding public key made available through a public key infrastructure.</p><p><strong>Trust.</strong> As outlined in Figure 2, when using a TEE, trust is delegated to the hardware manufacturer in most cases, that is Intel for Intel SGX/TDX or AMD for AMD SEV.</p><h3>Secure Multi-Party Computation</h3><p>In the previous paragraph, we saw how a TEE could be leveraged to perform private computation within a single helper party on data collected by an AdTech company.</p><p><em>Secure multi-party computation</em> (MPC) stands for another candidate to perform such private computation. In contrast to a TEE, MPC does not resort to a specific piece of hardware but only uses cryptographic protocols to ensure private computation. In addition, it leverages collaboration between several helper parties instead of resorting to a single one.</p><p><strong>Private computation.</strong><em> Private computation</em> is guaranteed by using a cryptographic technique called <em>secret sharing</em>. To be sure that data sent to helper parties will not be made available to them, this data (<em>i.e.</em> the secret) is first split into random pieces, called <em>secret</em> <em>shares,</em> such that we can reconstruct the original secret by aggregating them (<em>e.g.</em> via a sum). Only one of these random pieces is sent to each helper party; each of them being unable to reconstruct alone the initial data. Then, helper parties collaboratively compute the function of interest (<em>e.g.</em> a sum) by performing local computations on their random pieces of data and exchanging intermediate results between them. <br>In the IPA proposal, the AdTech provider is supposed to perform such secret sharing on encrypted reports before sending associated shares to helper parties.</p><p><strong>Correctness.</strong> MPC protocols do not meet the aforementioned <em>correctness </em>definition but are a relaxed version of it. More precisely, MPC protocols are meant to output the correct result of a given function. However, there is no way to guarantee, by default, that the function evaluated by the helper parties is the prescribed one. To meet <em>correctness</em>, MPC protocols could be run inside a TEE to benefit from the remote attestation feature, but the costs of these two PETs will then sum up.</p><p><strong>Threat Models &amp; Trust.</strong> Delegating computation to several helper parties naturally raises the following questions:</p><ul><li>What if helper parties collude to reconstruct my data?</li><li>What if a subset of the helper parties deviates from the prescribed protocol?</li></ul><p>Regarding the first question, collusion between parties is a potential risk that must be taken into account since it impairs <em>private computation</em> property. There is no efficient way to cope with this problem except managing to find helper parties in which we can place our trust.</p><p>On the other hand, the second question brings forward the assumption that some helper parties might be <em>malicious </em>and aim at breaking the prescribed protocol to recover plaintext data from intermediate computations on secret shares. In the scenario described in the IPA proposal, malicious helper parties are assumed to be in minority (1 out of 3). Under this assumption, MPC protocols can be enhanced by so-called <em>message authentification codes</em> (MACs) to yield correct results; at the cost of additional multiplications to be performed.</p><p><strong>Example.</strong> Figure 3 illustrates how to perform a simple private SUM query of two plaintext (for the sake of simplicity) integers, 8 and 15, using a simple instance of secret sharing. For each of these two numbers, the AdTech provider generates three random shares such that their sum equals the initial integer. A subset of these random shares is sent to each helper party, which computes intermediate sums. To obtain the final sum, all helper parties just sum up their intermediate sums to obtain the initial sum, which equals 8 + 15 = 23. Note that each helper party does not have access to the initial two integers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-HjZefO4cQvbR28WuYC5vw.png" /><figcaption>Figure 3. Illustration of a simple secret sharing scheme between three helper parties to compute a sum of two integers. The dice are used to refer to random numbers.</figcaption></figure><h3>Benchmark</h3><p>In this section, we provide a high-level benchmark between MPC and TEE with respect to key principles that should be verified when addressing <em>attribution</em> or <em>reporting use case</em>: accuracy, flexibility, and scalability. <br>In current browser vendors’ proposals, campaign optimisation via machine learning (ML) training is meant to be performed after noisy and aggregated data collection via a reporting API. In this benchmark, we also compare MPC and TEEs to perform campaign optimisation directly via helper parties. This use case could be envisioned in future versions of browser vendors’ proposals since it allows to use off-the-shelf ML algorithms that take as input non-aggregated features and labels.</p><h3>Accuracy</h3><p>In Table 1, we compare MPC and TEE from the accuracy lens. MPC protocols are based on two main primitives namely private addition and multiplication. <br>For simple reporting use cases, such as COUNT queries as proposed in the IPA proposal, we can leverage efficient private addition protocols to address them using MPC. For more advanced use cases such as campaign optimisation via machine learning (ML) training, we have to deal with non-linear functions that have to be approximated, in contrast to the use of a TEE. This approximation might lead to a performance drop for ML models.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*T6Q8LVHjVdmW7Faj8R3BKg.png" /><figcaption>Table 1. Comparison of MPC and TEE in terms of accuracy. Approximation of non-linear functions (e.g. <em>exp</em> or <em>log</em>) is usually performed via truncated Taylor expansions or polynomials.</figcaption></figure><h3>Flexibility</h3><p>In Table 2, we compare MPC and TEE from the flexibility lens; low flexibility meaning that associated algorithms may be cumbersome to design or often come with additional costs that must be paid.<br>Regarding the TEE, the use of an attestation mechanism to audit the code running inside it might lead to low flexibility despite meeting <em>correctness</em>. Indeed, if the browser vendor proposes to AdTech vendors the possibility to use custom SQL queries and ML training pipelines, then the former has to audit associated codes and verify that differential privacy is indeed met at the prescribed level. On the other hand, if the browser vendor proposes pre-defined templates to perform SQL queries or ML training to limit the code audit overhead, this yields a few degrees of freedom for AdTech to innovate.<br>For MPC, the lack of flexibility comes from (1) the use of additional protocols such as oblivious sorting to prevent information leakage by obfuscating the sorted location of any report, and (2) the necessity to provide efficient cryptographic primitives (e.g. private addition, private computation of the sigmoid function) to AdTech vendors to encompass a variety of private computations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sH2YZRbSA18N75Eno9XsJQ.png" /><figcaption>Table 2. Comparison of MPC and TEE in terms of flexibility. N/A means that there is no lack of flexibility.</figcaption></figure><h3>Scalability</h3><p>In Table 3, we compare MPC and TEE from the scalability lens. <br>Due to multi-party (but not independent in general) computations, MPC might suffer from large communication costs. These costs will be presented in thorough detail in a companion article focusing on MPC in practice for attribution, reporting, and campaign optimisation. In addition, because of the use of secret sharing, more local computations within each helper party is required (e.g. 2 sums instead of 1 sum in Figure 3).<br>Regarding TEEs, the computational overhead is mainly associated with encryption and memory access costs which vary from one application to another. They also depend on the TEE technology (e.g. AMD SEV or Intel SGX). More details will be provided in a companion article focusing on TEEs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DrPQ5dcItVvPagubEhan3A.png" /><figcaption>Table 3. Comparison of MPC and TEE in terms of scalability.</figcaption></figure><h3>What’s Next</h3><p>As pointed out in the Introduction, this article is part of a series of articles focused on the analysis of helper servers enhanced by MPC or TEE, for attribution and reporting use cases. Therefore, other articles will follow to complement this introductory content.</p><h3>References</h3><p>[1] <a href="https://en.wikipedia.org/wiki/Secure_multi-party_computation">Secure multi-party computation</a>; <a href="https://eprint.iacr.org/2020/300.pdf">https://eprint.iacr.org/2020/300.pdf</a></p><p>[2] <a href="https://en.wikipedia.org/wiki/Trusted_execution_environment">Trusted execution environment</a>; <a href="https://ieeexplore.ieee.org/document/7345265">https://ieeexplore.ieee.org/document/7345265</a></p><p>[3] <a href="https://github.com/patcg/docs-and-reports/tree/main/threat-model">docs-and-reports/threat-model at main · patcg/docs-and-reports</a>, <em>Introduction</em> section.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7e8cde13eca3" width="1" height="1" alt=""><hr><p><a href="https://medium.com/criteo-engineering/an-introduction-to-pets-for-attribution-and-reporting-7e8cde13eca3">An Introduction to PETs for Attribution and Reporting</a> was originally published in <a href="https://medium.com/criteo-engineering">Criteo Tech Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>