Jefferson's Wheel

Engineering and Technology Magazine (a publication of the British Institution of Engineering and Technology) has an article that highlights adversarial machine learning research: Artificial intelligence: the new ghost in the machine, 10 October 2018, by Chris Edwards.

Although researchers such as David Evans of the University of Virginia see a full explanation being a little way off in the future, the massive number of parameters encoded by DNNs and the avoidance of overtraining due to SGD may have an answer to why the networks can hallucinate images and, as a result, see things that are not there and ignore those that are.
…
He points to work by PhD student Mainuddin Jonas that shows how adversarial examples can push the output away from what we would see as the correct answer. “It could be just one layer [that makes the mistake]. But from our experience it seems more gradual. It seems many of the layers are being exploited, each one just a little bit. The biggest differences may not be apparent until the very last layer.”
…
Researchers such as Evans predict a lengthy arms race in attacks and countermeasures that may on the way reveal a lot more about the nature of machine learning and its relationship with reality.

I gave a keynote talk at USENIX Workshop of Offensive Technologies, Baltimore, Maryland, 13 August 2018.

The title and abstract are what I provided for the WOOT program, but unfortunately (or maybe fortunately for humanity!) I wasn’t able to actually figure out a talk to match the title and abstract I provided.

The history of security includes a long series of arms races, where a new technology emerges and is subsequently developed and exploited by both defenders and attackers. Over the past few years, “Artificial Intelligence” has re-emerged as a potentially transformative technology, and deep learning in particular has produced a barrage of amazing results. We are in the very early stages of understanding the potential of this technology in security, but more worryingly, seeing how it may be exploited by malicious individuals and powerful organizations. In this talk, I’ll look at what lessons might be learned from previous security arms races, consider how asymmetries in AI may be exploited by attackers and defenders, touch on some recent work in adversarial machine learning, and hopefully help progress-loving Luddites figure out how to survive in a world overrun by AI doppelgängers, GAN gangs, and gibbon-impersonating pandas.

I gave a keynote talk at the 1st Deep Learning and Security Workshop (co-located with the 39th IEEE Symposium on Security and Privacy). San Francisco, California. 24 May 2018

Over the past few years, there has been an explosion of research in security of machine learning and on adversarial examples in particular. Although this is in many ways a new and immature research area, the general problem of adversarial examples has been a core problem in information security for thousands of years. In this talk, I’ll look at some of the long-forgotten lessons from that quest and attempt to understand what, if anything, has changed now we are in the era of deep learning classifiers. I will survey the prevailing definitions for “adversarial examples”, argue that those definitions are unlikely to be the right ones, and raise questions about whether those definitions are leading us astray.

I spoke on Lessons from the Last 3000 Years of Adversarial Examples at Huawei’s Strategy and Technology Workshop in Shenzhen, China, 15 May 2018.

We also got to tour Huawei’s new research and development campus, under construction about 40 minutes from Shenzhen. It is pretty close to Disneyland, with its own railroad and villages themed after different European cities (Paris, Bologna, etc.).

Huawei’s New Research and Development Campus [More Pictures]

Unfortunately, pictures were not allowed on our tour of the production line. Not so surprising that nearly all of the work was done by machines, but was surprising to me how much of the human work left is completely robotic. The human workers (called “operators”) are mostly scanning QR codes on parts, and following the directions that light up with they do, or scanning bins and following directions on a screen to collect parts from bins and scanning them when they are put into the bin. This is the kind of system that leads to remarkably high production quality. The parts are mostly delivered on tapes that are fed into the machines, and many machines along the line are primarily for testing. There is a “bottleneck” marker that is placed on any points that are holding up the production line.

The public (at least to the factory) “grapey board” keeps track of the happiness of the workers — each operator puts up a smiley (or frowny) face on the board to show their mood for the day, monitored carefully by the managers. There is a batch of grapes to show performance for the month. If an operator does something good, a grape is colored green; if they do something bad, a grape is colored black. There was quite a bit of discussion among the people on the tour (mostly US and European-based professors) if such a management approach would be a good idea for our research groups… (or for department chairs for their faculty!)

In front of Huawei’s “White House”, with Battista Biggio [More Pictures]

Weilin Xu presented Feature Squeezing: Detecting Adversarial Examples in Deep Neural Networks at the Network and Distributed System Security Symposium 2018. San Diego, CA. 21 February 2018.

Paper: Weilin Xu, David Evans, Yanjun Qi. Feature Squeezing: Detecting Adversarial Examples in Deep Neural Networks. NDSS 2018. [PDF]

Project Site

I gave a talk at CISPA in Saarbrücken, Germany, on our work with Weilin Xu and Yanjun Qi on Adversarial Machine Learning: Are We Playing the Wrong Game?.

Machine learning classifiers are increasingly popular for security applications, and often achieve outstanding performance in testing. When deployed, however, classifiers can be thwarted by motivated adversaries who adaptively construct adversarial examples that exploit flaws in the classifier’s model. Much work on adversarial examples has focused on finding small distortions to inputs that fool a classifier. Previous defenses have been both ineffective and very expensive in practice. In this talk, I’ll describe a new very simple strategy, feature squeezing, that can be used to harden classifiers by detecting adversarial examples. Feature squeezing reduces the search space available to an adversary by coalescing samples that correspond to many different inputs in the original space into a single sample. Adversarial examples can be detected by comparing the model’s predictions on the original and squeezed sample. In practice, of course, adversaries are not limited to small distortions in a particular metric space. Indeed, in security applications like malware detection it may be possible to make large changes to an input without disrupting its intended malicious behavior. I’ll report on an evolutionary framework we have developed to search for such adversarial examples that can automatically find evasive variants against state-of-the-art classifiers. This suggests that work on adversarial machine learning needs a better definition of adversarial examples, and to make progress towards understanding how classifiers and oracles perceive samples differently.

I gave a talk at ECRYPT NET: Workshop on Crypto for the Cloud & Implementation (which was combined with Paris Crypto Day) on our group’s work on secure multi-party computation, using Bargav Jayaraman and Hannah Li‘s recent work on decentralizing certificate authorities as a motivating application.

I gave a talk at Berkeley’s International Computer Science Institute on Adversarial Machine Learning: Are We Playing the Wrong Game? (8 June 2017), focusing on the work Weilin Xu has been doing (in collaboration with myself and Yanjun Qi) on adversarial machine learning.

Abstract

Machine learning classifiers are increasingly popular for security applications, and often achieve outstanding performance in testing. When deployed, however, classifiers can be thwarted by motivated adversaries who adaptively construct adversarial examples that exploit flaws in the classifier’s model. Much work on adversarial examples, including Carlini and Wagner’s attacks which are the best results to date, has focused on finding small distortions to inputs that fool a classifier. Previous defenses have been both ineffective and very expensive in practice. In this talk, I’ll describe a new very simple strategy, feature squeezing, that can be used to harden classifiers by detecting adversarial examples. Feature squeezing reduces the search space available to an adversary by coalescing samples that correspond to many different inputs in the original space into a single sample. Adversarial examples can be detected by comparing the model’s predictions on the original and squeezed sample. In practice, of course, adversaries are not limited to small distortions in a particular metric space. Indeed, it may be possible to make large changes to an input without losing its intended malicious behavior. We have developed an evolutionary framework to search for such adversarial examples, and demonstrated that it can automatically find evasive variants against state-of-the-art classifiers. This suggests that work on adversarial machine learning needs a better definition of adversarial examples, and to make progress towards understanding how classifiers and oracles perceive samples differently.

Great to meet up with Wahooglers Adrienne Porter Felt, Ben Kreuter, Jonathan McCune, Samee Zahur (Google’s latest addition from my group), and (honorary UVAer interning at Google this summer) Riley Spahn at Google’s Research Summit on Security and Privacy this week in Mountain View.

As part of the meeting, the academic attendees were given a chance to give a 3-minute pitch to tell Google what we want them to do. The slides I used are below, but probably don’t make much sense by themselves.

The main modest proposal I tried to make is that Google should take it on as their responsibility to make sure nothing bad ever happens to anyone anywhere. They can start with nothing bad ever happening on the Internet, but with the Internet pretty much everywhere, should expand the scope to cover everywhere soon.

To start with an analogy from the days when Microsoft ruled computing. There was a time when Windows bluescreens were a frequent experience for most Windows users (and at the time, this pretty much mean all computer users). Microsoft analyzed the crashes and concluded that nearly all were because of bugs in device drivers, so it wasn’t their fault and was horribly unfair for them to be blamed for the crashes. Of course, to people losing their work because of a crash, it doesn’t really matter who’s code was to blame. By the end of the 90s, though, Microsoft took on the mission of reducing the problems with device drivers, and a lot of great work came out of this (e.g., the Static Driver Verifier), with dramatic improvements on the typical end user’s computing experience.

Today, Google rules a large chunk of computing. Lots of bad things happen on the Internet that are not Google’s fault. As the latest example in the news, the leaked NSA report of Russian attacks on election officials describes a phishing attack that exploits vulnerabilities in Microsoft Word. Its easy to put the blame on overworked election officials who didn’t pay enough attention to books on universal computation they read when they were children, or to put it on Microsoft for allowing Word to be exploited.

But, Google’s name is also all over this report – the emails when through gmail accounts, the attacks phished for Google credentials, and the attackers used plausibly-named gmail accounts. Even if Google isn’t too blame for the problems that enable such an attack, they are uniquely positioned to solve it, both because of their engineering capabilities and resources, but also because of the comprehensive view they have of what happens on the Internet and powerful ability to influence it.

Google is a big company, with lots of decentralized teams, some of which definitely seem to get this already. (I’d point to the work the Chrome Security Team has done, MOAR TLS, and RAPPOR as just a few of many examples of things that involve a mix of techincal and engineering depth and a broad mission to make computing better for everyone, not obviously connected to direct business interests.) But, there are also lots of places where Google doesn’t seem to be putting serious efforts into solving problems they could but viewing them as outside scope because its really someone else’s fault (my particular motivating example was PDF malware). As a company, Google is too capable, important, and ubiquitous to view problems as out-of-scope just because they are obviously undecidable or obviously really someone else’s fault.

[Also on Google +]

The video for my Enigma 2017 talk, “Classifiers under Attack” is now posted:

The talk focuses on work with Weilin Xu and Yanjun Qi on automatically evading malware classifiers using techniques from genetic programming. (See EvadeML.org for more details and links to code and papers, although some of the work I talked about at Enigma has not yet been published.)

Enigma was an amazing conference – one of the most worthwhile, and definitely the most diverse security/privacy conference I’ve been to in my career, both in terms of where people were coming from (nearly exactly 50% from industry and 50% from academic/government/non-profits), intellectual variety (range of talks from systems and crypto to neuroscience, law, and journalism), and the demographics of the attendees and speakers (not to mention a way-cool stage setup).

The model of having speakers do on-line practice talks with their session was also very valuable (Enigma requires speakers to agree to do three on-line practice talks sessions before the conference, and from what I hear most speakers and sessions did cooperate with this, and it showed in the quality of the sessions) and something I hope other conference will be able to adopt. You actually end up with talks that fit with each other, build of things others present, and avoid unnecessary duplication, as well as, improving all the talks by themselves.