<?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 Dr. Hari Koduvely on Medium]]></title>
        <description><![CDATA[Stories by Dr. Hari Koduvely on Medium]]></description>
        <link>https://medium.com/@hari.koduvely?source=rss-8995a096323e------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*6nWN2Cuuez5W5LydzXMwBg.jpeg</url>
            <title>Stories by Dr. Hari Koduvely on Medium</title>
            <link>https://medium.com/@hari.koduvely?source=rss-8995a096323e------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 23:00:41 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@hari.koduvely/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[Bayesian Research in NeurIPS2018]]></title>
            <link>https://medium.datadriveninvestor.com/bayesian-research-in-neurips2018-319cdbca71e9?source=rss-8995a096323e------2</link>
            <guid isPermaLink="false">https://medium.com/p/319cdbca71e9</guid>
            <category><![CDATA[bayesian-machine-learning]]></category>
            <category><![CDATA[nips]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[bayesian-statistics]]></category>
            <dc:creator><![CDATA[Dr. Hari Koduvely]]></dc:creator>
            <pubDate>Sun, 16 Dec 2018 12:38:05 GMT</pubDate>
            <atom:updated>2018-12-16T13:10:24.638Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*I-SllFUMChAVJJzMU4gG1A.jpeg" /></figure><h3>Bayesian Research at NeurIPS2018</h3><p>Neural Information Processing Systems conference (NeurIPS or NIPS previously) is considered as the most reputed academic conference in Artificial Intelligence. This year’s conference, <a href="https://nips.cc/Conferences/2018">NeurIPS2018</a> was held in the beautiful city of Montreal from 3rd to 8th December. Thanks to Facebook, the conference was live streamed and videos of all lectures is available on <a href="https://www.facebook.com/nipsfoundation/">NIPS Foundation Facebook page</a>. There were some very interesting keynote lectures. The most fascinating talk I found was <a href="https://nips.cc/Conferences/2018/Schedule?showEvent=12487">Bioelectric Computation Outside the Nervous System, Primitive Cognition and Synthetic Morphology</a> by Professor Micheal Levin. Other noticeable keynote lectures were <a href="https://nips.cc/Conferences/2018/Schedule?showEvent=12486">Reproducible, Reusable and Robust Reinforcement Learning</a> by Professor Joelle Pineau, <a href="https://nips.cc/Conferences/2018/Schedule?showEvent=12346">Making Algorithms Trustworthy</a> by David Spiegelhalter and <a href="https://nips.cc/Conferences/2018/Schedule?showEvent=12469">Designing Computer Systems for Software 2.0</a> by Kunle Olukotun. There was also a very interesting workshop on <a href="https://nips.cc/Conferences/2018/Schedule?showEvent=10923">Machine Learning for Molecules and Materials</a> for discussing applications in Physical Sciences.</p><p>I was following NeurIPS2018 mainly for understanding latest research going on in <strong>Bayesian Inference</strong> and its use in Machine Learning and Artificial Intelligence. As expected there were quite a few papers, more than 70, spanning the 6 day conference. Day 0 tutorial on <a href="https://nips.cc/Conferences/2018/Schedule?showEvent=10984">Scalable Bayesian Inference</a> by Professor David Dunson was a very useful one. This tutorial contains an overview of state-of-the-art approaches for analyzing massive data sets using Bayesian statistical methods. Professor Dunson discussed simple approaches for scaling up commonly used Markov chain Monte Carlo (MCMC) algorithms. Some examples discussed were embarrassingly parallel (EP) MCMC, approximate MCMC, stochastic approximation, hybrid optimization &amp; sampling, and modularization. These methods have applications in areas such as computational advertising, genomics and neurosciences.</p><p>There were also two workshops related to <strong>Bayesian Inference</strong>. One was on <a href="https://nips.cc/Conferences/2018/Schedule?showEvent=10906">Bayesian Deep Learning</a> and the other was on <a href="https://nips.cc/Conferences/2018/Schedule?showEvent=10905">Non-Parametric Bayesian Inference</a> both are very active areas of research currently.</p><p>The oral presentations and posters covered several aspects of <strong>Bayesian Inference</strong>, both theoretical advancements and applications in machine learning. The topics covered includes Bayesian Deep Learning, Bayesian Reinforcement Learning, Bayesian Optimization, Variational Inference, Variational Auto-Encoder, Markov Chain Monte Carlo (MCMC) Methods, Representation Learning or Meta Learning, Cognitive Science, Differential Privacy, Approximate Bayesian Methods and Bayesian Networks. This wide set of topics shows the importance of Bayesian methods in Machine Learning and Artificial Intelligence.</p><p>I will here summarize some of the talks/posters I found interesting, full list of talks/posters can be found on <a href="https://papers.nips.cc/book/advances-in-neural-information-processing-systems-31-2018">NeurIPS2018 Proceedings</a> page.</p><p><a href="https://nips.cc/Conferences/2018/Schedule?showEvent=12618"><strong>Neural Architecture Search with Bayesian Optimization and Optimal Transport</strong></a><strong><br></strong>In this work, the authors develop NASBOT, a Gaussian process based Bayesian Optimization framework for neural architecture search. The authors develop a distance metric in the space of neural network architectures, which can be computed efficiently via an optimal transport program.</p><p><a href="https://nips.cc/Conferences/2018/Schedule?showEvent=11310"><strong>Constructing Deep Neural Networks by Bayesian Network Structure Learning</strong></a> <br>In this paper the authors introduce a principled approach for unsupervised structure learning of deep neural networks. They propose a new interpretation for depth and inter-layer connectivity, where conditional independencies in the input distribution are encoded hierarchically in the network structure, which enables the depth of the network to be determined inherently. The proposed method casts the problem of neural network structure learning as a problem of Bayesian network structure learning.</p><p><a href="https://nips.cc/Conferences/2018/Schedule?showEvent=11445"><strong>Explaining Deep Learning Models — A Bayesian Non-parametric Approach</strong></a><strong><br></strong>In this work, the authors propose a novel technical approach that augments a Bayesian non-parametric regression mixture model with multiple elastic nets. Using the enhanced mixture model, one can extract generalizable insights for a target model through a global approximation.</p><p><a href="https://nips.cc/Conferences/2018/Schedule?showEvent=11664"><strong>Bayesian Adversarial Learning</strong></a><strong><br></strong>Deep neural networks are vulnerable to adversarial attacks and the standard defensive approaches is through formulating it as a robust optimization problem. Here one minimizes a point estimate of the worst-case loss derived from an <em>adversary data-generating distribution</em>. In this work, a novel robust training framework called <em>Bayesian Robust Learning</em>, is proposed, in which a distribution is put on the adversarial data-generating distribution to account for the uncertainty of the adversarial data-generating process.</p><p><a href="https://nips.cc/Conferences/2018/Schedule?showEvent=11617"><strong>Bayesian Distributed Stochastic Gradient Descent</strong></a><strong><br></strong>This work discusses a high-throughput algorithm for training deep neural networks on parallel clusters. This algorithm uses amortized inference in a deep generative model, to perform joint posterior predictive inference of mini-batch gradient computation times, in a compute cluster specific manner. Particularly, this algorithm mitigates the straggler effect in synchronous, gradient-based optimization by choosing an optimal cutoff beyond which mini-batch gradient messages from slow workers are ignored.</p><p><a href="https://nips.cc/Conferences/2018/Schedule?showEvent=12594"><strong>Bayesian Model-Agnostic Meta-Learning</strong></a><strong><br></strong>In this paper, the authors propose a novel Bayesian model-agnostic meta-learning method for learning from a small data sets. The proposed method combines efficient gradient-based meta-learning with nonparametric variational inference in a principled probabilistic framework.</p><p><a href="https://nips.cc/Conferences/2018/Schedule?showEvent=11405"><strong>Beauty-in-averageness and its contextual modulations: A Bayesian statistical account</strong></a><strong><br></strong>Understanding how humans perceive the likability of high-dimensional objects’’ such as faces is an important problem in both cognitive science and AI/ML. It is well known from psychology literature that human assessment of facial attractiveness is context-dependent. In this paper, the authors hypothesize that human preference for an object is increased when it incurs lower encoding cost, in particular when its perceived <em>statistical typicality</em> is high, in consonance with Barlow’s seminal efficient coding hypothesis.</p><p><a href="https://nips.cc/Conferences/2018/Schedule?showEvent=11379"><strong>Bayesian Inference of Temporal Task Specifications from Demonstrations</strong></a><strong><br></strong>When observing task demonstrations, human apprentices are able to identify whether a given task is executed correctly long before they gain expertise in actually performing that task. The authors present <em>Bayesian Specification Inference</em>, a probabilistic model for inferring task specification as a temporal logic formula. Authors incorporate methods from probabilistic programming to define their priors, along with a domain-independent likelihood function to enable sampling-based inference.</p><p><a href="https://nips.cc/Conferences/2018/Schedule?showEvent=11971"><strong>Predictive Approximate Bayesian Computation via Saddle Point</strong></a><strong><br></strong>Approximate Bayesian computation (ABC) is an important methodology for Bayesian inference when the likelihood function is intractable. In this paper, the authors introduce an optimization-based ABC framework that addresses deficiencies in existing methods. Leveraging a generative model for posterior and joint distribution matching, the authors show that ABC can be framed as saddle point problems, whose objectives can be accessed directly with samples.</p><p><a href="https://nips.cc/Conferences/2018/Schedule?showEvent=11906"><strong>Reinforcement Learning with Multiple Experts: A Bayesian Model Combination Approach</strong></a><strong><br></strong> In this paper, the authors apply Bayesian Model Combination with multiple experts in a way that learns to trust a good combination of experts as training progresses.</p><p><a href="https://nips.cc/Conferences/2018/Schedule?showEvent=11786"><strong>Variational Bayesian Monte Carlo</strong></a><strong><br></strong>Many probabilistic models of interest in scientific computing and machine learning are intractable and require access to the gradient or a large number of likelihood evaluations. The authors introduce here a novel sample-efficient inference framework, <em>Variational Bayesian Monte Carlo</em> (VBMC). VBMC combines variational inference with Gaussian-process based, active-sampling Bayesian quadrature, using the latter to efficiently approximate the intractable integral in the variational objective.</p><p><a href="https://nips.cc/Conferences/2018/Schedule?showEvent=11722"><strong>Inference in Deep Gaussian Processes using Stochastic Gradient Hamiltonian Monte Carlo</strong></a><strong><br></strong>Deep Gaussian Processes (DGPs) are hierarchical generalizations of Gaussian Processes that combine well calibrated uncertainty estimates with the high flexibility of multilayer models. One of the biggest challenges with these models is that exact inference is intractable. In this work, the authors provide evidence for the non-Gaussian nature of the posterior and they apply the Stochastic Gradient Hamiltonian Monte Carlo method to generate samples from posterior distribution.</p><p><a href="https://nips.cc/Conferences/2018/Schedule?showEvent=11534"><strong>Algorithmic Assurance: An Active Approach to Algorithmic Testing using Bayesian Optimization</strong></a><strong><br></strong>In this work, the authors introduce algorithmic assurance, the problem of testing whether machine learning algorithms are conforming to their intended design goal. The authors mathematically formulate this task as an optimization problem of an expensive, black-box function. They use an active learning approach based on Bayesian optimization to solve this optimization problem.</p><p><a href="https://nips.cc/Conferences/2018/Schedule?showEvent=12713"><strong>Discretely Relaxing Continuous Variables for tractable Variational Inference</strong></a><strong> <br></strong>The authors explore a new research direction in Bayesian variational inference with discrete latent variable priors where they exploit Kronecker matrix algebra for efficient and exact computations of the evidence lower bound (ELBO). This results in posterior samples consist of sparse and low-precision quantized integers which permit fast inference on hardware limited devices.</p><p><a href="https://nips.cc/Conferences/2018/Schedule?showEvent=11256"><strong>Wasserstein Variational Inference</strong></a><strong><br></strong>This paper introduces <em>Wasserstein Variational Inference</em>, a new form of approximate Bayesian inference based on optimal transport theory. Wasserstein variational inference uses a new family of divergences that includes both f-divergences and the Wasserstein distance as special cases. This technique results in a very stable likelihood-free training method that can be used with implicit distributions and probabilistic programs.</p><p><a href="https://nips.cc/Conferences/2018/Schedule?showEvent=11623"><strong>Learning Latent Subspaces in Variational Autoencoders</strong></a><strong><br></strong>It is often difficult to interpret the latent space representations learned using <em>Variational Autoencoders</em> (VAE). The authors propose a VAE-based generative model, which is capable of extracting features correlated to binary labels in the data and structuring it in a latent subspace, which is easy to interpret.</p><p><a href="https://nips.cc/Conferences/2018/Schedule?showEvent=11748"><strong>Constrained Graph Variational Autoencoders for Molecule Design</strong></a><strong><br></strong>With an emphasis on applications in chemistry, the authors explore the task of learning to generate graphs that conform to a distribution observed in training data. They propose a Variational Autoencoder model in which both encoder and decoder are graph-structured. They show that by using appropriate shaping of the latent space, this model allows to design molecules that are (locally) optimal in desired properties.</p><p>I might have omitted here several other important work on <strong>Bayesian Inference</strong> presented at NeurIPS2018. This is partly because of my personal choices and also I might not have noticed them among the several hundreds of papers and posters presented at the conference. I would encourage the readers to search through the <a href="https://papers.nips.cc/book/advances-in-neural-information-processing-systems-31-2018">NeurIPS2018 Proceedings</a> to find other interesting papers.</p><p>In summary the extensive list of work on Bayesian Inference presented at NeurIPS2018 shows the relevance of the subject in the modern era of Machine Learning and Artificial Intelligence. There has been some important advances last year in the state-of-art Bayesian Inference in terms of better algorithms for posterior density estimation, and application in problems ranging from interpretation of Deep Learning models to design of new molecules.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=319cdbca71e9" width="1" height="1" alt=""><hr><p><a href="https://medium.datadriveninvestor.com/bayesian-research-in-neurips2018-319cdbca71e9">Bayesian Research in NeurIPS2018</a> was originally published in <a href="https://medium.datadriveninvestor.com/">DataDrivenInvestor</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Bayesian Learning helps in Decentralization of AI]]></title>
            <link>https://medium.com/@hari.koduvely/bayesian-learning-helps-in-decentralization-of-ai-da6a4bb76f7a?source=rss-8995a096323e------2</link>
            <guid isPermaLink="false">https://medium.com/p/da6a4bb76f7a</guid>
            <category><![CDATA[behavioral-biometrics]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[bayesian-machine-learning]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Dr. Hari Koduvely]]></dc:creator>
            <pubDate>Tue, 11 Sep 2018 03:07:42 GMT</pubDate>
            <atom:updated>2018-09-11T03:07:42.766Z</atom:updated>
            <content:encoded><![CDATA[<p>Decentralization is the new mantra in AI. Due to the explosive growth of mobile phones and other personal devices it has become necessary to deploy AI agents on small devices to provide better user experiences and protect sensitive personal data. However there are many challenges in achieving this goal. I will address two of the main challenges here and discuss how AI research is progressing towards finding a solution.</p><p>The first challenge is that Deep Neural Networks, the state of art machine learning models that power most of today’s AI agents, are computationally intensive to run on small devices. The second challenge, which is related to the first, is that it requires order of thousands of data points to train Deep Neural Networks which is typically not available on small devices.</p><p>Research to solve both these problems finds that Bayesian Inference, which is a branch of AI evolved from Bayesian Statistics, is a promising approach [1]. In a recent blogpost Qualcomm researchers have mentioned the advantages of using Bayesian Deep Learning for power efficient AI on mobile chips [2]. Currently this is demonstrated mainly for the computer vision domain applications such as classifying images and understanding scenes. This article describes some details of this method and how it can be generalized to other applications such as Behavioral Biometrics where <a href="https://www.zighra.com/">Zighra</a> is focussing. Our vision is to make Behavioral Biometrics based authentication entirely on device to protect privacy of users.</p><p>Let us look at how Bayesian inference can solve the first problem in this post. I will discuss how Bayesian methods can help in learning from small datasets in the second part of this blog post. The size of a DNN can be reduced during the training phase using a technique called Deep Compression [3]. This consists of 3 separate steps namely Pruning, Quantization of Weights and Huffman Coding. Pruning involves forcing the network to learn only the important connections. This reduces the number of connections typically by a factor of 10. During Quantization, weights are discretized which enables weight sharing between Neurons and helps to reduce the number of bits needed to store each connection. Huffman Coding is a lossless data compression algorithm where variable length codes are assigned to input characters of a string such that length of the codes are inversely proportional to frequencies of characters. Moreover the codes are from a family called Prefix Codes where no code is a prefix of another code in the sequence and the sequence is completely decodable. It was demonstrated that overall this approach can reduce the size of a network up to 50 times. In general the idea is to induce sparsity in the network architecture during the training phase of DNN which helps in avoiding overfitting of data through regularization, compression of the network and acceleration of the training itself.</p><p>Bayesian Inference provides a principled framework for training sparse DNN models such that the degree of sparsity can be tuned individually to each neuron or layer in a network. This approach was originated from the Sparse Bayesian Learning framework introduced by Tipping more than a decade ago [4]. To understand this better we need to get into some technical details of Bayesian inference.</p><p>Let D be a dataset containing N rows, each consisting of tuples xn,yn, where n=1,…,N. Goal is to predict value of y given a new observation x using the probability distribution p(y|x,w). Here w is parameter characterizing the distribution. In Bayesian approach one starts with a prior distribution p(w) for the parameter w which represents our prior knowledge or belief about the phenomena that we are modelling. After making some observations about the phenomena through data D, one update the belief through the well-known Bayesian theorem</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/235/1*_HeWsnGCRppSXV4INfzwIA.png" /></figure><p>This process of estimating the posterior distribution from prior after receiving some data is called <em>Bayesian Inference</em>. This is not as simple as it sounds since usually it is very difficult to obtain the distribution p(D) analytically and one has to use approximate methods.</p><p>Standard approach to find posterior distribution is Markov Chain Monte Carlo Simulation aka MCMC. Recently another approach, called Variational Inference, is becoming very popular. Here the posterior distribution p(w|D) is approximated by a parametric distribution q(w). The optimal value of the variational parameter is found by minimizing the Lagrangian functional</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/365/1*DkEQo7zlJXUbjiSQGBMvtg.png" /></figure><p>Here the first term on right hand side is the average value of log likelihood function taken over the distribution q(w) given by the equation</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/324/1*xv3Im7RN1Ur8D3YONZTQhA.png" /></figure><p>The sum is over the N data points in D. Second term is the Kullback-Leibler (KL) divergence, which is a measure of how good is the approximation q(w) to p(w|D). The minimization can be done very efficiently and in a scalable manner using a stochastic variational inference algorithm developed by Kingma and Welling [5]</p><p>So how one can use Bayesian Inference to reduce the size of Deep Neural Networks? It was shown by Geoffrey Hinton and team a few years back that if one multiplies weights of a neural network by a continuous noise N(1, alpha=p/1-p), where p is a parameter called <em>drop out rate</em>, it leads to effective regularization of the network and avoiding of overfitting [6]. The physical meaning of drop out is that during training some weights are set to zero with probability p. In the limit of p -&gt; 1this implies that some weights are permanently set to zero or in other words removed from the network and hence a reduction in the network size. It can be shown that this can be cast as a Bayesian Inference problem with a specific prior distribution for the weights called <em>improper log-uniform prior </em>given by</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/159/1*By0qKtKakzjSoIcMEZVF9w.png" /></figure><p>where wij is the weight of the link connecting Neurons i and j. Using this approach researchers have achieved close to 300 times reduction in the number of parameters on popular DNN architectures such as LeNet with insignificant reduction in accuracy [7]</p><p>These research findings are quite encouraging for implementing DNN models for behavioral biometrics based authentication completely on a device. The first step here is to demonstrate that DNNs are useful in processing time series data originating from smartphone sensors such as accelerometer and gyroscope and capable of identifying a user uniquely based on their usage of smartphones. Next step is to compress these DNN architectures using Bayesian compression scheme as explained in this post. Zighra AI team is working towards this direction to make AI completely decentralized and on-device for the protection of privacy of users.</p><h3>References</h3><ol><li>Bayesian Compression for Deep Learning, C. Louizos et.al., Proceedings of Neural Information Processing Systems Conference (NIPS), 2017</li><li>How Algorithmic Advances Make Power Efficient AI Possible <a href="https://www.qualcomm.com/news/onq/2018/08/13/how-algorithmic-advances-make-power-efficient-ai-possible">https://www.qualcomm.com/news/onq/2018/08/13/how-algorithmic-advances-make-power-efficient-ai-possible</a></li><li>Deep Compression: Compressing Deep Neural Networks with pruning, training quantization and huffman coding. S. Han et.al., International Conference on Learning Representations (ICLR), 2016.</li><li>Sparse Bayesian Learning and Relevance Vector Machine, ME Tipping, Journal of Machine Learning Research, 211–244, 2001.</li><li>Auto Encoding Variational Bayes, D. P. Kingma et.al., International Conference on Learning Representations (ICLR), 2014.</li><li>Dropout: A Simple Way to Prevent Neural Networks from Overfitting, Journal of Machine Learning Research, 1929−1958, 2014</li><li>Variational Dropouts Sparsifies Deep Neural Networks, International Conference on Machine Learning (ICML), 2017</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=da6a4bb76f7a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Learning Representations from a Cybersecurity Perspective]]></title>
            <link>https://medium.com/@hari.koduvely/learning-representations-from-a-cybersecurity-perspective-406b37668297?source=rss-8995a096323e------2</link>
            <guid isPermaLink="false">https://medium.com/p/406b37668297</guid>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[adversarial-network]]></category>
            <category><![CDATA[differential-privacy]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Dr. Hari Koduvely]]></dc:creator>
            <pubDate>Sat, 19 May 2018 02:15:37 GMT</pubDate>
            <atom:updated>2018-05-19T02:15:37.174Z</atom:updated>
            <content:encoded><![CDATA[<p>The sixth <em>International Conference on</em> <em>Learning Representations</em> (<a href="https://iclr.cc/">ICLR 2018</a>)was held in the beautiful city of Vancouver in the first week of May 2018. Most of the papers presented in this conference were related to learning representations using Deep Neural Networks and similar approaches. There were close to 350 presentations/posters. It would be impossible to summarize all the new research presented in ICLR 2018 in a blog. You can find a summary of the 3 papers which got the Best Paper awards in this <a href="https://medium.com/syncedreview/iclr-2018s-best-papers-variant-adam-spherical-cnns-and-meta-learning-6b48dca83e8b">medium post</a>.</p><p>Although did not attend the conference, I was watching it closely from a Cybersecurity perspective. I was interested in the latest research on representation learning for identity, privacy and threat detection. Since ICLR is not a mainstream security conference, there were not many papers directly related to Cybersecurity. In this blog post I will try to categorize and summarize all the research work presented in ICLR 2018, which are relevant for Cybersecurity. I hope this post would be useful for all Cybersecurity Researchers.</p><p>The oral and poster presentations relevant for Cybersecurity can be grouped into 4 categories. The first one deals with <a href="https://en.wikipedia.org/wiki/Differential_privacy"><em>Differential Privacy</em></a>, a term used for describing the privacy loss that results to individuals when their private information is used for creating a data driven product or service. In the recent context of allegations of personal data on Facebook being misused, I don’t need to stress the importance of this research topic. The second group is about <a href="https://en.wikipedia.org/wiki/Adversarial_machine_learning"><em>Adversarial Learning</em></a>. Here the focus is on identifying and preventing ways in which a malicious adversary could compromise a machine learning algorithm by carefully crafting the input data. This is another very important and emerging area of machine learning, because more and more products and services are incorporating machine learning and artificial intelligence these days. Knowledge about adversarial learning would become important to secure these systems fully. The third group of research work is about <a href="https://en.wikipedia.org/wiki/Malware"><em>Malware Detection</em></a><em> </em>using Deep Learning. The final category is that of <em>Anomaly Detection</em>.</p><p><strong>I. Differential Privacy</strong></p><ol><li><a href="https://openreview.net/forum?id=rkZB1XbRZ">Scalable Private Learning with PATE</a> <br>This work was mainly contributed by researchers at the <a href="https://ai.google/research/teams/brain">Google Brain</a> team. It builds on a method published last year by the same researchers known as <a href="https://arxiv.org/abs/1610.05755"><em>Private Aggregation of Teacher Ensembles or PATE</em></a>. The idea is to first train a machine learning model using data from disjoint set of users. Then use these machine learning model as Teachers to train a Student machine learning model <em>using the noisy voting among all of the teachers </em>as the supervised signal. By choosing a carefully-calibrated <a href="https://www.coursera.org/learn/data-results/lecture/SqNET/adding-laplacian-noise-and-proving-differential-privacy">Laplacian Noise</a> for this purpose and publishing only the Student model, it can be shown that this scheme provides rigorous differential privacy guarantees.</li><li><a href="https://openreview.net/forum?id=r17_wzJPM">SGD on Random Mixtures: Private Machine Learning Under Data Breach Threats</a> <br>Here the authors of the paper proposes a simple method called <em>Stochastic Gradient Descent on Random Mixtures (SGDRM)</em> to provide Differential Privacy. Their algorithm constructs several random noisy mixtures of the original dataset and performs normal SGD on these mixtures. For the cases of least square and linear neural network, the authors shows that their scheme provides differential privacy guarantees and convergence guarantee. They also gives empirical evidence that SGDRM works for the case of general neural networks.</li><li><a href="https://openreview.net/forum?id=BJ0hF1Z0b">Learning Differentially Private Recurrent Language Models</a><br>This is another paper from the <a href="https://ai.google/research/teams/brain">Google Brain</a> team. In this work the authors shows the feasibility of training large Recurrent Neural Network Language Models (LSTM) with user-level differential privacy guarantees, yet with negligible compromise in the accuracy of predictions. It uses the concept of <a href="https://ai.googleblog.com/2017/04/federated-learning-collaborative.html">Federated Machine Learning</a> where a <em>Decentralised Dataset</em> is used for training the machine learning models.</li></ol><p><strong>II. Adversarial Learning</strong></p><ol><li><a href="https://openreview.net/forum?id=rJzIBfZAb">Towards Deep Learning Models Resistant to Adversarial Attacks</a><br>This research work from MIT proposes <a href="https://en.wikipedia.org/wiki/Robust_optimization">Robust Optimization</a> approach to train Deep Neural Networks against adversarial attacks. The authors claims that their approach can provide concrete security guarantee that would protect against a well-defined class of adversaries. Also the approach seems to be reliable and universal.</li><li><a href="https://openreview.net/forum?id=BkJ3ibb0-">Defense GAN-Protecting Classifiers Against Adversarial Attacks Using Generative Models</a><br>Deep Neural Network predictions are susceptible to adversarial perturbations of the input data. A small, carefully designed but un noticeable, perturbation in the input data can cause the DNN to predict wrong class labels. The authors, from University of Maryland, have proposed a novel Generative Adversarial Network called Defense-GAN which can solve this problem. At the time of inference, Defense-GAN finds a close output to a given image which does not contain the adversarial changes. This output is then fed to the DNN classifier. Their method can be used with any classifier and not just DNN.</li><li><a href="https://openreview.net/forum?id=S10qYwywf">Combating Adversarial Attacks Using Sparse Representations</a><br>This research work from University of California, Santa Barbara proposes a different approach to solve the same problem mentioned above. For linear classifiers, the authors show that a sparsifying front end is very effective against certain type of attacks. They then extend this concept to DNNs, showing that a “locally linear” model can be used to develop a theoretical foundation for crafting attacks and defenses.</li><li><a href="https://openreview.net/forum?id=Bys4ob-Rb">Certified Defenses Against Adversarial Examples</a><br>This paper from Stanford University proposes a yet another approach to solve the same problem. This approach is based on a relatively new area of optimization called <a href="https://en.wikipedia.org/wiki/Semidefinite_programming"><em>Semidefinite Programming</em></a>. It gives an <em>upper bound</em> <em>on the performance of classifier</em> on an adversarially perturbed input sample. This upper bound is also differentiable, and therefore can be jointly optimized along with the network parameters. This would make it an adaptive regularizer that defense against all type of attacks.</li><li><a href="https://openreview.net/forum?id=SyZI0GWCZ">Decision-Based Adversarial Attacks: Reliable Attacks Against Black-Box Machine Learning Models</a><br>Where as the previous papers discusses the adversarial learning in a more laboratory setting, where the detailed model information and class probabilities of the underlaying DNN is known for constructing the adversarial perturbation, this paper addresses the same problem in a real world context, for example autonomous cars.</li></ol><p>There are a few more similar research work reported for adversarial learning in the papers mentioned below.</p><ol><li><a href="https://openreview.net/forum?id=SJ9nVSkvG">Black-box Attacks on Deep Neural Networks Via Gradient Estimation</a></li><li><a href="https://openreview.net/forum?id=Sk9yuql0Z">Mitigating Adversarial Effects through Randomization</a></li><li><a href="https://openreview.net/forum?id=BkeqO7x0-">Unsupervised Cipher Cracking Using Discrete GANs</a></li></ol><p>In general what I find is that though there is considerable research happening in the area of adversarial learning, most of them are focussed on Deep Neural Networks and Computer Vision problems, where this problem has shown to be very critical. However from the Cybersecurity point of view, relevance of adversarial attacks needs to be probed further, and research works need to take into account other learning methods, which are commonly used in this domain.</p><p><strong>III. Malware Detection</strong></p><ol><li><a href="https://openreview.net/forum?id=HkHrmM1PM">Deep Convolutional Malware Classifiers Can Learn from Raw Executables and Labels Only</a><br>This paper is from the maker of popular free anti-virus software Avast, in collaboration with Czech Academy of Sciences. The authors have developed a Convolutional DNN to detect malicious <em>Portable Executables</em> on Windows machines, by learning from their raw sequences of bytes and labels without any domain-specific feature extraction or preprocessing. They show that on a dataset of 20 million unpacked half megabyte Portable Executables, CNN achieves performance almost on par with the traditional machine learning algorithms based on handcrafted features of Avast.</li><li><a href="https://openreview.net/forum?id=rkjatuyvM">Monotonic Models for Real Time Dynamic Malware Detection</a> <br>This work originates from Kaspersky Labs and Samsung-HSE Laboratory in Russia. The authors have developed a method of real-time malware detection using a program’s execution time. The basic assumption here is that the probability of being malicious increases monotonically as new lines of codes are added into the execution log, whereas probability of being a benign program would remain the same. This feature makes it suitable for identifying malwares in real time.</li></ol><p><strong>IV. ANOMALY DETECTION</strong></p><ol><li><a href="https://openreview.net/forum?id=BJJLHbb0-">Deep Autoencoding Gaussian Mixture Model for Unsupervised Anomaly Detection</a><br>This paper from NEC Laboratories and Washington State University addresses some of the problems with the state-of-art conventional anomaly detection methods. The authors then proposes a novel Deep Autoencoding Gaussian Mixture Model (DAGMM) as a solution for unsupervised anomaly detection.</li></ol><p>In summary ICLR 2018 shows that Deep Learning and related machine learning techniques continue to be the main focus area of research today for representation learning. More research needs to be done addressing representation learning that is directly relevant for the Cybersecurity domain.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=406b37668297" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>