Entries Tagged "artificial intelligence"
Page 3 of 4
DARPA is launching a program aimed at vulnerability discovery via human-assisted AI. The new DARPA program is called CHESS (Computers and Humans Exploring Software Security), and they’re holding a proposers day in a week and a half.
This is the kind of thing that can dramatically change the offense/defense balance.
Artificial intelligence technologies have the potential to upend the longstanding advantage that attack has over defense on the Internet. This has to do with the relative strengths and weaknesses of people and computers, how those all interplay in Internet security, and where AI technologies might change things.
You can divide Internet security tasks into two sets: what humans do well and what computers do well. Traditionally, computers excel at speed, scale, and scope. They can launch attacks in milliseconds and infect millions of computers. They can scan computer code to look for particular kinds of vulnerabilities, and data packets to identify particular kinds of attacks.
Humans, conversely, excel at thinking and reasoning. They can look at the data and distinguish a real attack from a false alarm, understand the attack as it’s happening, and respond to it. They can find new sorts of vulnerabilities in systems. Humans are creative and adaptive, and can understand context.
Computers — so far, at least — are bad at what humans do well. They’re not creative or adaptive. They don’t understand context. They can behave irrationally because of those things.
Humans are slow, and get bored at repetitive tasks. They’re terrible at big data analysis. They use cognitive shortcuts, and can only keep a few data points in their head at a time. They can also behave irrationally because of those things.
AI will allow computers to take over Internet security tasks from humans, and then do them faster and at scale. Here are possible AI capabilities:
- Discovering new vulnerabilities — and, more importantly, new types of vulnerabilities in systems, both by the offense to exploit and by the defense to patch, and then automatically exploiting or patching them.
- Reacting and adapting to an adversary’s actions, again both on the offense and defense sides. This includes reasoning about those actions and what they mean in the context of the attack and the environment.
- Abstracting lessons from individual incidents, generalizing them across systems and networks, and applying those lessons to increase attack and defense effectiveness elsewhere.
- Identifying strategic and tactical trends from large datasets and using those trends to adapt attack and defense tactics.
That’s an incomplete list. I don’t think anyone can predict what AI technologies will be capable of. But it’s not unreasonable to look at what humans do today and imagine a future where AIs are doing the same things, only at computer speeds, scale, and scope.
Both attack and defense will benefit from AI technologies, but I believe that AI has the capability to tip the scales more toward defense. There will be better offensive and defensive AI techniques. But here’s the thing: defense is currently in a worse position than offense precisely because of the human components. Present-day attacks pit the relative advantages of computers and humans against the relative weaknesses of computers and humans. Computers moving into what are traditionally human areas will rebalance that equation.
Roy Amara famously said that we overestimate the short-term effects of new technologies, but underestimate their long-term effects. AI is notoriously hard to predict, so many of the details I speculate about are likely to be wrong — and AI is likely to introduce new asymmetries that we can’t foresee. But AI is the most promising technology I’ve seen for bringing defense up to par with offense. For Internet security, that will change everything.
This essay previously appeared in the March/April 2018 issue of IEEE Security & Privacy.
Researchers found that they could confuse the road sign detection algorithms of self-driving cars by adding stickers to the signs on the road. They could, for example, cause a car to think that a stop sign is a 45 mph speed limit sign. The changes are subtle, though — look at the photo from the article.
“Robust Physical-World Attacks on Machine Learning Models,” by Ivan Evtimov, Kevin Eykholt, Earlence Fernandes, Tadayoshi Kohno, Bo Li, Atul Prakash, Amir Rahmati, and Dawn Song:
Abstract: Deep neural network-based classifiers are known to be vulnerable to adversarial examples that can fool them into misclassifying their input through the addition of small-magnitude perturbations. However, recent studies have demonstrated that such adversarial examples are not very effective in the physical world–they either completely fail to cause misclassification or only work in restricted cases where a relatively complex image is perturbed and printed on paper. In this paper we propose a new attack algorithm–Robust Physical Perturbations (RP2)– that generates perturbations by taking images under different conditions into account. Our algorithm can create spatially-constrained perturbations that mimic vandalism or art to reduce the likelihood of detection by a casual observer. We show that adversarial examples generated by RP2 achieve high success rates under various conditions for real road sign recognition by using an evaluation methodology that captures physical world conditions. We physically realized and evaluated two attacks, one that causes a Stop sign to be misclassified as a Speed Limit sign in 100% of the testing conditions, and one that causes a Right Turn sign to be misclassified as either a Stop or Added Lane sign in 100% of the testing conditions.
The objective of this CRA is to perform enabling basic and applied research to extend the reach, situational awareness, and operational effectiveness of large heterogeneous teams of intelligent systems and Soldiers against dynamic threats in complex and contested environments and provide technical and operational superiority through fast, intelligent, resilient and collaborative behaviors. To achieve this, ARL is requesting proposals that address three key Research Areas (RAs):
RA1: Distributed Intelligence: Establish the theoretical foundations of multi-faceted distributed networked intelligent systems combining autonomous agents, sensors, tactical super-computing, knowledge bases in the tactical cloud, and human experts to acquire and apply knowledge to affect and inform decisions of the collective team.
RA2: Heterogeneous Group Control: Develop theory and algorithms for control of large autonomous teams with varying levels of heterogeneity and modularity across sensing, computing, platforms, and degree of autonomy.
RA3: Adaptive and Resilient Behaviors: Develop theory and experimental methods for heterogeneous teams to carry out tasks under the dynamic and varying conditions in the physical world.
And while we’re on the subject, this is an excellent report on AI and national security.
This article argues that AI technologies will make image, audio, and video forgeries much easier in the future.
Combined, the trajectory of cheap, high-quality media forgeries is worrying. At the current pace of progress, it may be as little as two or three years before realistic audio forgeries are good enough to fool the untrained ear, and only five or 10 years before forgeries can fool at least some types of forensic analysis. When tools for producing fake video perform at higher quality than today’s CGI and are simultaneously available to untrained amateurs, these forgeries might comprise a large part of the information ecosystem. The growth in this technology will transform the meaning of evidence and truth in domains across journalism, government communications, testimony in criminal justice, and, of course, national security.
I am not worried about fooling the “untrained ear,” and more worried about fooling forensic analysis. But there’s an arms race here. Recording technologies will get more sophisticated, too, making their outputs harder to forge. Still, I agree that the advantage will go to the forgers and not the forgery detectors.
Last month at the RSA Conference, I saw a lot of companies selling security incident response automation. Their promise was to replace people with computers – sometimes with the addition of machine learning or other artificial intelligence techniques – and to respond to attacks at computer speeds.
While this is a laudable goal, there’s a fundamental problem with doing this in the short term. You can only automate what you’re certain about, and there is still an enormous amount of uncertainty in cybersecurity. Automation has its place in incident response, but the focus needs to be on making the people effective, not on replacing them — security orchestration, not automation.
This isn’t just a choice of words – it’s a difference in philosophy. The US military went through this in the 1990s. What was called the Revolution in Military Affairs (RMA) was supposed to change how warfare was fought. Satellites, drones and battlefield sensors were supposed to give commanders unprecedented information about what was going on, while networked soldiers and weaponry would enable troops to coordinate to a degree never before possible. In short, the traditional fog of war would be replaced by perfect information, providing certainty instead of uncertainty. They, too, believed certainty would fuel automation and, in many circumstances, allow technology to replace people.
Of course, it didn’t work out that way. The US learned in Afghanistan and Iraq that there are a lot of holes in both its collection and coordination systems. Drones have their place, but they can’t replace ground troops. The advances from the RMA brought with them some enormous advantages, especially against militaries that didn’t have access to the same technologies, but never resulted in certainty. Uncertainty still rules the battlefield, and soldiers on the ground are still the only effective way to control a region of territory.
But along the way, we learned a lot about how the feeling of certainty affects military thinking. Last month, I attended a lecture on the topic by H.R. McMaster. This was before he became President Trump’s national security advisor-designate. Then, he was the director of the Army Capabilities Integration Center. His lecture touched on many topics, but at one point he talked about the failure of the RMA. He confirmed that military strategists mistakenly believed that data would give them certainty. But he took this change in thinking further, outlining the ways this belief in certainty had repercussions in how military strategists thought about modern conflict.
McMaster’s observations are directly relevant to Internet security incident response. We too have been led to believe that data will give us certainty, and we are making the same mistakes that the military did in the 1990s. In a world of uncertainty, there’s a premium on understanding, because commanders need to figure out what’s going on. In a world of certainty, knowing what’s going on becomes a simple matter of data collection.
I see this same fallacy in Internet security. Many companies exhibiting at the RSA Conference promised to collect and display more data and that the data will reveal everything. This simply isn’t true. Data does not equal information, and information does not equal understanding. We need data, but we also must prioritize understanding the data we have over collecting ever more data. Much like the problems with bulk surveillance, the “collect it all” approach provides minimal value over collecting the specific data that’s useful.
In a world of uncertainty, the focus is on execution. In a world of certainty, the focus is on planning. I see this manifesting in Internet security as well. My own Resilient Systems – now part of IBM Security – allows incident response teams to manage security incidents and intrusions. While the tool is useful for planning and testing, its real focus is always on execution.
Uncertainty demands initiative, while certainty demands synchronization. Here, again, we are heading too far down the wrong path. The purpose of all incident response tools should be to make the human responders more effective. They need both the ability and the capability to exercise it effectively.
When things are uncertain, you want your systems to be decentralized. When things are certain, centralization is more important. Good incident response teams know that decentralization goes hand in hand with initiative. And finally, a world of uncertainty prioritizes command, while a world of certainty prioritizes control. Again, effective incident response teams know this, and effective managers aren’t scared to release and delegate control.
Like the US military, we in the incident response field have shifted too much into the world of certainty. We have prioritized data collection, preplanning, synchronization, centralization and control. You can see it in the way people talk about the future of Internet security, and you can see it in the products and services offered on the show floor of the RSA Conference.
Automation, too, is fixed. Incident response needs to be dynamic and agile, because you are never certain and there is an adaptive, malicious adversary on the other end. You need a response system that has human controls and can modify itself on the fly. Automation just doesn’t allow a system to do that to the extent that’s needed in today’s environment. Just as the military shifted from trying to replace the soldier to making the best soldier possible, we need to do the same.
For some time, I have been talking about incident response in terms of OODA loops. This is a way of thinking about real-time adversarial relationships, originally developed for airplane dogfights, but much more broadly applicable. OODA stands for observe-orient-decide-act, and it’s what people responding to a cybersecurity incident do constantly, over and over again. We need tools that augment each of those four steps. These tools need to operate in a world of uncertainty, where there is never enough data to know everything that is going on. We need to prioritize understanding, execution, initiative, decentralization and command.
At the same time, we’re going to have to make all of this scale. If anything, the most seductive promise of a world of certainty and automation is that it allows defense to scale. The problem is that we’re not there yet. We can automate and scale parts of IT security, such as antivirus, automatic patching and firewall management, but we can’t yet scale incident response. We still need people. And we need to understand what can be automated and what can’t be.
The word I prefer is orchestration. Security orchestration represents the union of people, process and technology. It’s computer automation where it works, and human coordination where that’s necessary. It’s networked systems giving people understanding and capabilities for execution. It’s making those on the front lines of incident response the most effective they can be, instead of trying to replace them. It’s the best approach we have for cyberdefense.
Automation has its place. If you think about the product categories where it has worked, they’re all areas where we have pretty strong certainty. Automation works in antivirus, firewalls, patch management and authentication systems. None of them is perfect, but all those systems are right almost all the time, and we’ve developed ancillary systems to deal with it when they’re wrong.
Automation fails in incident response because there’s too much uncertainty. Actions can be automated once the people understand what’s going on, but people are still required. For example, IBM’s Watson for Cyber Security provides insights for incident response teams based on its ability to ingest and find patterns in an enormous amount of freeform data. It does not attempt a level of understanding necessary to take people out of the equation.
From within an orchestration model, automation can be incredibly powerful. But it’s the human-centric orchestration model – the dashboards, the reports, the collaboration – that makes automation work. Otherwise, you’re blindly trusting the machine. And when an uncertain process is automated, the results can be dangerous.
Technology continues to advance, and this is all a changing target. Eventually, computers will become intelligent enough to replace people at real-time incident response. My guess, though, is that computers are not going to get there by collecting enough data to be certain. More likely, they’ll develop the ability to exhibit understanding and operate in a world of uncertainty. That’s a much harder goal.
Yes, today, this is all science fiction. But it’s not stupid science fiction, and it might become reality during the lifetimes of our children. Until then, we need people in the loop. Orchestration is a way to achieve that.
This essay previously appeared on the Security Intelligence blog.
Interesting research: “Using Artificial Intelligence to Identify State Secrets,” by Renato Rocha Souza, Flavio Codeco Coelho, Rohan Shah, and Matthew Connelly.
Abstract: Whether officials can be trusted to protect national security information has become a matter of great public controversy, reigniting a long-standing debate about the scope and nature of official secrecy. The declassification of millions of electronic records has made it possible to analyze these issues with greater rigor and precision. Using machine-learning methods, we examined nearly a million State Department cables from the 1970s to identify features of records that are more likely to be classified, such as international negotiations, military operations, and high-level communications. Even with incomplete data, algorithms can use such features to identify 90% of classified cables with <11% false positives. But our results also show that there are longstanding problems in the identification of sensitive information. Error analysis reveals many examples of both overclassification and underclassification. This indicates both the need for research on inter-coder reliability among officials as to what constitutes classified material and the opportunity to develop recommender systems to better manage both classification and declassification.
This is some interesting research. You can fool facial recognition systems by wearing glasses printed with elements of other people’s faces.
Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, and Michael K. Reiter, “Accessorize to a Crime: Real and Stealthy Attacks on State-of-the-Art Face Recognition“:
ABSTRACT: Machine learning is enabling a myriad innovations, including new algorithms for cancer diagnosis and self-driving cars. The broad use of machine learning makes it important to understand the extent to which machine-learning algorithms are subject to attack, particularly when used in applications where physical security or safety is at risk. In this paper, we focus on facial biometric systems, which are widely used in surveillance and access control. We define and investigate a novel class of attacks: attacks that are physically realizable and inconspicuous, and allow an attacker to evade recognition or impersonate another individual. We develop a systematic method to automatically generate such attacks, which are realized through printing a pair of eyeglass frames. When worn by the attacker whose image is supplied to a state-of-the-art face-recognition algorithm, the eyeglasses allow her to evade being recognized or to impersonate another individual. Our investigation focuses on white-box face-recognition systems, but we also demonstrate how similar techniques can be used in black-box scenarios, as well as to avoid face detection.
In their experiment, computers were able to make their own form of encryption using machine learning, without being taught specific cryptographic algorithms. The encryption was very basic, especially compared to our current human-designed systems. Even so, it is still an interesting step for neural nets, which the authors state “are generally not meant to be great at cryptography:.
This story is more about AI and neural networks than it is about cryptography. The algorithm isn’t any good, but is a perfect example of what I’ve heard called “Schneier’s Law“: Anyone can design a cipher that they themselves cannot break.
Research paper. Note that the researchers work at Google.
Sidebar photo of Bruce Schneier by Joe MacInnis.