Interesting research: “Facial Misrecognition Systems: Simple Weight Manipulations Force DNNs to Err Only on Specific Persons“:
Abstract: In this paper we describe how to plant novel types of backdoors in any facial recognition model based on the popular architecture of deep Siamese neural networks, by mathematically changing a small fraction of its weights (i.e., without using any additional training or optimization). These backdoors force the system to err only on specific persons which are preselected by the attacker. For example, we show how such a backdoored system can take any two images of a particular person and decide that they represent different persons (an anonymity attack), or take any two images of a particular pair of persons and decide that they represent the same person (a confusion attack), with almost no effect on the correctness of its decisions for other persons. Uniquely, we show that multiple backdoors can be independently installed by multiple attackers who may not be aware of each other’s existence with almost no interference.
We have experimentally verified the attacks on a FaceNet-based facial recognition system, which achieves SOTA accuracy on the standard LFW dataset of 99.35%. When we tried to individually anonymize ten celebrities, the network failed to recognize two of their images as being the same person in 96.97% to 98.29% of the time. When we tried to confuse between the extremely different looking Morgan Freeman and Scarlett Johansson, for example, their images were declared to be the same person in 91.51% of the time. For each type of backdoor, we sequentially installed multiple backdoors with minimal effect on the performance of each one (for example, anonymizing all ten celebrities on the same model reduced the success rate for each celebrity by no more than 0.91%). In all of our experiments, the benign accuracy of the network on other persons was degraded by no more than 0.48% (and in most cases, it remained above 99.30%).
It’s a weird attack. On the one hand, the attacker has access to the internals of the facial recognition system. On the other hand, this is a novel attack in that it manipulates internal weights to achieve a specific outcome. Given that we have no idea how those weights work, it’s an important result.
Posted on February 3, 2023 at 7:07 AM •
Mandiant is reporting on a trojaned Windows installer that targets Ukrainian users. The installer was left on various torrent sites, presumably ensnaring people downloading pirated copies of the operating system:
Mandiant uncovered a socially engineered supply chain operation focused on Ukrainian government entities that leveraged trojanized ISO files masquerading as legitimate Windows 10 Operating System installers. The trojanized ISOs were hosted on Ukrainian- and Russian-language torrent file sharing sites. Upon installation of the compromised software, the malware gathers information on the compromised system and exfiltrates it. At a subset of victims, additional tools are deployed to enable further intelligence gathering. In some instances, we discovered additional payloads that were likely deployed following initial reconnaissance including the STOWAWAY, BEACON, and SPAREPART backdoors.
One obvious solution would be for Microsoft to give the Ukrainians Windows licenses, so they don’t have to get their software from sketchy torrent sites.
Posted on December 20, 2022 at 7:30 AM •
Interesting research: “ImpNet: Imperceptible and blackbox-undetectable backdoors in compiled neural networks, by Tim Clifford, Ilia Shumailov, Yiren Zhao, Ross Anderson, and Robert Mullins:
Abstract: Early backdoor attacks against machine learning set off an arms race in attack and defence development. Defences have since appeared demonstrating some ability to detect backdoors in models or even remove them. These defences work by inspecting the training data, the model, or the integrity of the training procedure. In this work, we show that backdoors can be added during compilation, circumventing any safeguards in the data preparation and model training stages. As an illustration, the attacker can insert weight-based backdoors during the hardware compilation step that will not be detected by any training or data-preparation process. Next, we demonstrate that some backdoors, such as ImpNet, can only be reliably detected at the stage where they are inserted and removing them anywhere else presents a significant challenge. We conclude that machine-learning model security requires assurance of provenance along the entire technical pipeline, including the data, model architecture, compiler, and hardware specification.
Ross Anderson explains the significance:
The trick is for the compiler to recognise what sort of model it’s compiling—whether it’s processing images or text, for example—and then devising trigger mechanisms for such models that are sufficiently covert and general. The takeaway message is that for a machine-learning model to be trustworthy, you need to assure the provenance of the whole chain: the model itself, the software tools used to compile it, the training data, the order in which the data are batched and presented—in short, everything.
Posted on October 11, 2022 at 7:18 AM •
What makes Symbiote different from other Linux malware that we usually come across, is that it needs to infect other running processes to inflict damage on infected machines. Instead of being a standalone executable file that is run to infect a machine, it is a shared object (SO) library that is loaded into all running processes using LD_PRELOAD (T1574.006), and parasitically infects the machine. Once it has infected all the running processes, it provides the threat actor with rootkit functionality, the ability to harvest credentials, and remote access capability.
Researchers have unearthed a discovery that doesn’t occur all that often in the realm of malware: a mature, never-before-seen Linux backdoor that uses novel evasion techniques to conceal its presence on infected servers, in some cases even with a forensic investigation.
No public attribution yet.
So far, there’s no evidence of infections in the wild, only malware samples found online. It’s unlikely this malware is widely active at the moment, but with stealth this robust, how can we be sure?
Posted on June 22, 2022 at 6:07 AM •
Mandiant is reporting on a new botnet.
The group, which security firm Mandiant is calling UNC3524, has spent the past 18 months burrowing into victims’ networks with unusual stealth. In cases where the group is ejected, it wastes no time reinfecting the victim environment and picking up where things left off. There are many keys to its stealth, including:
- The use of a unique backdoor Mandiant calls Quietexit, which runs on load balancers, wireless access point controllers, and other types of IoT devices that don’t support antivirus or endpoint detection. This makes detection through traditional means difficult.
- Customized versions of the backdoor that use file names and creation dates that are similar to legitimate files used on a specific infected device.
- A live-off-the-land approach that favors common Windows programming interfaces and tools over custom code with the goal of leaving as light a footprint as possible.
- An unusual way a second-stage backdoor connects to attacker-controlled infrastructure by, in essence, acting as a TLS-encrypted server that proxies data through the SOCKS protocol.
Unpacking this threat group is difficult. From outward appearances, their focus on corporate transactions suggests a financial interest. But UNC3524’s high-caliber tradecraft, proficiency with sophisticated IoT botnets, and ability to remain undetected for so long suggests something more.
Throughout their operations, the threat actor demonstrated sophisticated operational security that we see only a small number of threat actors demonstrate. The threat actor evaded detection by operating from devices in the victim environment’s blind spots, including servers running uncommon versions of Linux and network appliances running opaque OSes. These devices and appliances were running versions of operating systems that were unsupported by agent-based security tools, and often had an expected level of network traffic that allowed the attackers to blend in. The threat actor’s use of the QUIETEXIT tunneler allowed them to largely live off the land, without the need to bring in additional tools, further reducing the opportunity for detection. This allowed UNC3524 to remain undetected in victim environments for, in some cases, upwards of 18 months.
Posted on May 4, 2022 at 6:15 AM •
New paper: “Planting Undetectable Backdoors in Machine Learning Models“:
Abstract: Given the computational cost and technical expertise required to train machine learning models, users may delegate the task of learning to a service provider. We show how a malicious learner can plant an undetectable backdoor into a classifier. On the surface, such a backdoored classifier behaves normally, but in reality, the learner maintains a mechanism for changing the classification of any input, with only a slight perturbation. Importantly, without the appropriate “backdoor key”, the mechanism is hidden and cannot be detected by any computationally-bounded observer. We demonstrate two frameworks for planting undetectable backdoors, with incomparable guarantees.
First, we show how to plant a backdoor in any model, using digital signature schemes. The construction guarantees that given black-box access to the original model and the backdoored version, it is computationally infeasible to find even a single input where they differ. This property implies that the backdoored model has generalization error comparable with the original model. Second, we demonstrate how to insert undetectable backdoors in models trained using the Random Fourier Features (RFF) learning paradigm or in Random ReLU networks. In this construction, undetectability holds against powerful white-box distinguishers: given a complete description of the network and the training data, no efficient distinguisher can guess whether the model is “clean” or contains a backdoor.
Our construction of undetectable backdoors also sheds light on the related issue of robustness to adversarial examples. In particular, our construction can produce a classifier that is indistinguishable from an “adversarially robust” classifier, but where every input has an adversarial example! In summary, the existence of undetectable backdoors represent a significant theoretical roadblock to certifying adversarial robustness.
EDITED TO ADD (4/20): Cory Doctorow wrote about this as well.
Posted on April 19, 2022 at 3:12 PM •
I hope this is true:
According to Jens Zimmermann, the German coalition negotiations had made it “quite clear” that the incoming government of the Social Democrats (SPD), the Greens and the business-friendly liberal FDP would reject “the weakening of encryption, which is being attempted under the guise of the fight against child abuse” by the coalition partners.
Such regulations, which are already enshrined in the interim solution of the ePrivacy Regulation, for example, “diametrically contradict the character of the coalition agreement” because secure end-to-end encryption is guaranteed there, Zimmermann said.
Introducing backdoors would undermine this goal of the coalition agreement, he added.
I have written about this.
Posted on December 8, 2021 at 1:19 PM •
Even before Apple made its announcement, law enforcement shifted their battle for backdoors to client-side scanning. The idea is that they wouldn’t touch the cryptography, but instead eavesdrop on communications and systems before encryption or after decryption. It’s not a cryptographic backdoor, but it’s still a backdoor—and brings with it all the insecurities of a backdoor.
I’m part of a group of cryptographers that has just published a paper discussing the security risks of such a system. (It’s substantially the same group that wrote a similar paper about key escrow in 1997, and other “exceptional access” proposals in 2015. We seem to have to do this every decade or so.) In our paper, we examine both the efficacy of such a system and its potential security failures, and conclude that it’s a really bad idea.
We had been working on the paper well before Apple’s announcement. And while we do talk about Apple’s system, our focus is really on the idea in general.
Ross Anderson wrote a blog post on the paper. (It’s always great when Ross writes something. It means I don’t have to.) So did Susan Landau. And there’s press coverage in the New York Times, the Guardian, Computer Weekly, the Financial Times, Forbes, El Pais (English translation), NRK (English translation), and—this is the best article of them all—the Register. See also this analysis of the law and politics of client-side scanning from last year.
Posted on October 15, 2021 at 9:30 AM •
We knew the basics of this story, but it’s good to have more detail.
Here’s me in 2015 about this Juniper hack. Here’s me in 2007 on the NSA backdoor.
Posted on September 9, 2021 at 6:13 AM •
In this post, I’ll collect links on Apple’s iPhone backdoor for scanning CSAM images. Previous links are here and here.
Apple says that hash collisions in its CSAM detection system were expected, and not a concern. I’m not convinced that this secondary system was originally part of the design, since it wasn’t discussed in the original specification.
Good op-ed from a group of Princeton researchers who developed a similar system:
Our system could be easily repurposed for surveillance and censorship. The design wasn’t restricted to a specific category of content; a service could simply swap in any content-matching database, and the person using that service would be none the wiser.
EDITED TO ADD (8/30): Good essays by Matthew Green and Alex Stamos, Ross Anderson, Edward Snowden, and Susan Landau. And also Kurt Opsahl.
EDITED TO ADD (9/6): Apple is delaying implementation of the scheme.
Posted on August 20, 2021 at 8:54 AM •
Sidebar photo of Bruce Schneier by Joe MacInnis.