Security Guru Gary McGraw on What's Needed To Secure Machine Learning Apps
John K. Waters talks with McGraw about his new focus on securing the next generation of applications.
- By John K. Waters
It's been more than a decade since a group of software security mavens set out to create a "fact-based" set of best practices for developing and growing an enterprise-wide software security program, which came to be known as the Building Security In Maturity Model (BSIMM). It was the first maturity model for security initiatives created entirely from real-world data.
The ring leader of that group, cybersecurity guru Gary McGraw, has been pushing developers to take responsibility for building secure software for almost two decades. So, I wasn't surprised when I learned that he'd rounded up another posse to chase down the facts about the security of machine learning (ML) systems -- and in the process, developed a new set of recommendations for building security in.
"I was seeing all this breathless coverage of security in machine learning," he told me, "but it was all related to attacks. It reminded me of what we saw in software security 25 years ago, when people were barely talking about it. It was clear that AI and machine learning had moved into the mainstream really fast -- and that no one had done a real risk analysis at the architectural level. So, we did one," he commented in our interview.
That work led McGraw and his colleagues (Harold Figueroa, Victor Shepardson and Richie Bonett) to publish the first-ever risk framework to guide development of secure ML ("An Architectural Risk Analysis of Machine Learning Systems: Toward More Secure Machine Learning"), and to found the Berryville Institute of Machine Learning, a research think tank dedicated to safe, secure, and ethical development of AI technologies. (McGraw lives in Berryville, VA.)
The paper is a guide for developers, engineers, designers, and anyone creating applications and services that use ML technologies. It focuses on providing advice on building security into ML systems from a security engineering perspective -- which involves "understanding how ML systems are designed for security, teasing out possible security engineering risks, and making such risks explicit," the authors explained. "We are also interested in the impact of including an ML system as part of a larger design. Our basic motivating question is: how do we secure ML systems pro-actively while we are designing and building them? This architectural risk analysis (ARA) is an important first step in our mission to help engineers and researchers secure ML systems."
"We need to do better work to secure our ML systems," they added, "moving well beyond attack-of-the day and penetrate-and-patch towards real security engineering."
The paper, which is free for download, is a must-read. It includes a wealth of information and insights from experts in the field focused on the unique security risks of ML systems. It identifies 78 specific risks associated with a generic ML system, and provides an interactive ML risk framework.
It also offers the researchers' Top Ten ML Security Risks. "These risks come in two relatively distinct flavors," they wrote, "both equally valid: some are risks associated with the intentional actions of an attacker; others are risks associated with an intrinsic design flaw. Intrinsic design flaws emerge when engineers with good intentions screw things up. Of course, attackers can also go after intrinsic design flaws complicating the situation."
The list includes:
Probably the most commonly discussed attacks against machine learning have come to be known as adversarial examples. The basic idea is to fool a machine learning system by providing malicious input often involving very small perturbations that cause the system to make a false prediction or categorization. Though coverage and resulting attention might be disproportionately large, swamping out other important ML risks, adversarial examples are very much real.
Data play an outsized role in the security of an ML system. That's because an ML system learns to do what it does directly from data. If an attacker can intentionally manipulate the data being used by an ML system in a coordinated fashion, the entire system can be compromised. Data poisoning attacks require special attention. In particular, ML engineers should consider what fraction of the training data an attacker can control and to what extent.
Online system manipulation
An ML system is said to be "online" when it continues to learn during operational use, modifying its behavior over time. In this case a clever attacker can nudge the still-learning system in the wrong direction on purpose through system input and slowly "retrain" the ML system to do the wrong thing. Note that such an attack can be both subtle and reasonably easy to carry out. This risk is complex, demanding that ML engineers consider data provenance, algorithm choice, and system operations in order to properly address it.
Transfer learning attack
In many cases in the real world, ML systems are constructed by taking advantage of an already-trained base model which is then fine-tuned to carry out a more specific task. A data transfer attack takes place when the base system is compromised (or otherwise unsuitable), making unanticipated behavior defined by the attacker possible
Data protection is difficult enough without throwing ML into the mix. One unique challenge in ML is protecting sensitive or confidential data that, through training, are built right into a model. Subtle but effective extraction attacks against an ML system's data are an important category of risk.
Because data play an outsize role in ML security, considering data provenance and integrity is essential. Are the data suitable and of high enough quality to support ML? Are sensors reliable? How is data integrity preserved? Understanding the nature of ML system data sources (both during training and during execution) is of critical importance. Data borne risks are particularly hairy when it comes to public data sources (which might be manipulated or poisoned) and online models.
When science and engineering are sloppy, everyone suffers. Unfortunately, because of its inherent inscrutability and the hyper-rapid growth of the field, ML system results are often under-reported, poorly described, and otherwise impossible to reproduce. When a system can't be reproduced and nobody notices, bad things can happen.
ML systems are often very powerful. Sometimes they can be too powerful for their own good. When an ML system "memorizes" its training data set, it will not generalize to new data, and is said to be overfitting. Overfit models are particularly easy to attack. Keep in mind that overfitting is possible in concert with online system manipulation and may happen while a system is running.
Data are often encoded, filtered, re-represented, and otherwise processed before use in an ML system (in most cases by a human engineering group). Encoding integrity issues can bias a model in interesting and disturbing ways. For example, encodings that include metadata may allow an ML model to "solve" a categorization problem by overemphasizing the metadata and ignoring the real categorization problem.
If an attacker can interpose between an ML system and the world, a direct attack on output may be possible. The inscrutability of ML operations (that is, not really understanding how they do what they do) may make an output integrity attack that much easier since an anomaly may be harder to spot.
McGraw has been preaching the gospel of security through better software at conferences, on college campuses, as an advisor to leading security companies, and in a long running podcast for nearly three decades. He wrote eight books on the subject, including the best-selling "White Hat/Black Hat Trilogy," which includes Building Secure Software: How to Avoid Security Problems the Right Way, which he penned with security expert John Viega, Exploiting Software: How to Break Code, which he wrote with the semi-legendary Greg Hoglund, and his solo effort, Software Security: Building Security In.
A next step for the Berryville gang would be to take the insights they've gleaned working with a generic ML model and apply it to specific ML systems, McGraw said. "The other thing you could do is to think about what controls can be associated with those risks," he added, "so that you can put in place controls to manage those risks appropriately. There's a ton of work still to be done here."
John K. Waters is the editor in chief of a number of Converge360.com sites, with a focus on high-end development, AI and future tech. He's been writing about cutting-edge technologies and culture of Silicon Valley for more than two decades, and he's written more than a dozen books. He also co-scripted the documentary film Silicon Valley: A 100 Year Renaissance, which aired on PBS. He can be reached at firstname.lastname@example.org.