Memory Re-orderings Achieve 95% Cross-Process Signal Inference As a Timerless Side-Channel

Modern computer processors routinely re-order memory operations to enhance performance, a practice enabled by relaxed memory models. Sean Siddens from Trail of Bits, Sanya Srivastava of the University of Washington, Reese Levine from Duke University, and et al. demonstrate a novel side-channel attack, termed Memory DisOrder, which exploits these re-orderings to detect activity in other processes running on the same chip. Their research reveals that a significant number of mainstream processors , including those from X86, Arm, and Apple , are vulnerable to signals leaking across process boundaries. This vulnerability allows for the implementation of attacks such as covert channels, achieving data transfer rates of up to 16 bits per second on Apple M3 GPUs, and precise application fingerprinting, identifying underlying deep neural network architectures. The team further discovered methods to amplify these re-orderings, potentially increasing covert channel bandwidth to nearly 30 kilobits per second on X86 CPUs, highlighting a fundamental security challenge in contemporary computing systems.

Memory DisOrder Attack Exploits Reordering Vulnerabilities

To enhance processing efficiency, modern CPUs and GPUs employ relaxed memory models, allowing memory operations to be re-ordered. This research demonstrates a novel side-channel attack, termed Memory DisOrder, that exploits these memory re-orderings to infer activity on other processes without reliance on timers or privileged access. To establish the presence of these vulnerabilities, the team engineered a comprehensive fuzz testing system comprised of two core processes: a Listener and a Stressor. The Listener executes litmus tests, repeatedly assessing for memory re-orderings, while the Stressor concurrently stresses the system to encourage more frequent and observable re-orderings. This approach allowed researchers to systematically probe processor behaviour under varying conditions, revealing cross-process signalling opportunities.

Scientists developed four distinct Listener frameworks, each employing unique techniques to detect re-orderings. The Basic Testing Framework utilises C++ threads and relaxed atomic memory accesses, enabling detection of all instruction re-orderings and fuzzing memory locations. The Litmus7 framework builds upon existing work, leveraging inline assembly and thread reuse to enhance throughput, while the Perpetual framework eliminates synchronization by storing algebraic sequences and analysing traces for re-ordering indicators. A GPU Parallel Testing Framework was also implemented, harnessing WebGPU to execute numerous litmus tests concurrently across GPU threads, with parameters like workgroup size and memory padding subject to fuzzing. Complementing the Listener, the Stressor employed techniques to induce system stress, including mirroring prior work and a novel approach designed to maximise re-ordering frequency. Experiments revealed that the frequency of these re-orderings increases when other cores are active, suggesting hardware optimizations are a key factor. This vulnerability allows for the implementation of covert channels and application fingerprinting techniques.

The team measured a covert channel achieving up to 16 bits per second with 95% accuracy on an Apple M3 GPU, demonstrating a viable method for data exfiltration. Further investigation on X86 CPUs showed the potential to amplify this signal, achieving nearly 30,000 bits per second. Tests prove that the system can reliably fingerprint Deep Neural Network (DNN) architectures in a closed-world scenario on both CPUs and the Apple M3 GPU. Specifically, the research team observed unique patterns of re-ordering frequencies for different DNNs, including resnet50, googlenet, vgg16, mobilenetv3, and alexnet, allowing for accurate classification of the running architecture.

Researchers employed a Message Passing litmus test, a short sequence of memory operations, to detect re-orderings, observing whether a read operation occurs before a corresponding write operation, indicating a re-ordering. By repeatedly executing this test, the attacker process can detect patterns of system pressure created by the victim process, revealing information about the victim’s activity. This work highlights the low capability of Memory DisOrder, requiring only two threads and shared memory, without the need for timers or specialized hardware access.

The research team conducted a fuzzing campaign, utilizing a ‘Stressor’ process to induce system load and a ‘Listener’ process to monitor memory re-orderings, confirming the presence of a detectable signal across processes. Data shows that the vulnerability crosses virtualization boundaries, functioning even within KVM on Linux, further expanding its potential impact. Researchers demonstrated the vulnerability across a broad range of devices, including CPUs from Arm, X86, and Apple, as well as GPUs from NVIDIA, AMD, and Apple. Through a fuzzing campaign, they established that these re-orderings are consistently observable and can be leveraged to infer activity on other processes running on the same system. Furthermore, the team showed that manipulating low-level system details can significantly amplify re-orderings, increasing the potential data transfer rate of a covert channel to nearly 30 kilobytes per second on X86 CPUs. The authors acknowledge that identifying the precise causes of these memory re-orderings remains a challenge, and future research focusing on specific processors could lead to more refined attacks and, crucially, more effective mitigation strategies.