Exploiting Congestion in Multi-GPU Interconnects for Covert and Side Channel Attacks

Core Concepts

Adversaries can exploit congestion on high-speed multi-GPU interconnects like NVLink to launch covert and side channel attacks, enabling them to infer private information about victims' activities without special permissions.

Abstract

The content explores two types of attacks on multi-GPU systems:

Cross-GPU Covert Channel Attack:

The attacker and victim are positioned on separate GPUs connected via NVLink.
The attacker can covertly transmit a message by modulating the data transfer rates on the shared NVLink, which leads to differentiable profiling times.
The authors demonstrate a covert channel attack with a bandwidth of 45.5 kbps and a low error rate of 3.22%.

Cross-GPU Side Channel Attack:

A background spy application continuously monitors the NVLink data transfer latency caused by other users' activities.
The NVLink traffic traces are analyzed to correlate with the victim's actions, allowing the attacker to infer the applications the victim is running.
The authors demonstrate the attack by identifying specific HPC applications from the OpenMM benchmark based on the NVLink congestion side channel leakages.

The paper highlights the vulnerability of multi-GPU systems to these attacks and advocates for heightened awareness and security measures in implementing multi-GPU interconnects.

Stats

The covert channel attack achieved a bandwidth of 45.5 kbps with an error rate of 3.22%.

Quotes

"Leveraging this insight, we develop a covert channel attack across two GPUs with a bandwidth of 45.5 kbps and a low error rate, and introduce a side channel attack enabling attackers to fingerprint applications through the shared NVLink interconnect."

Key Insights Distilled From

Beyond the Bridge

by Yicheng Zhan... at arxiv.org 04-08-2024

https://arxiv.org/pdf/2404.03877.pdf

Deeper Inquiries

How can multi-GPU systems be designed or modified to mitigate the risks of these congestion-based attacks?

To mitigate the risks of congestion-based attacks on multi-GPU systems, several design or modification strategies can be implemented. One approach is to introduce traffic shaping mechanisms that regulate the data flow on the interconnects, preventing adversaries from exploiting congestion patterns for covert or side channel attacks. By implementing quality of service (QoS) policies, the system can prioritize critical data transfers and limit the impact of potential attacks. Additionally, enhancing encryption and authentication protocols for data exchanged over the interconnects can bolster security. By encrypting sensitive information and ensuring secure communication channels, the system can thwart eavesdropping attempts by malicious actors. Furthermore, implementing intrusion detection systems (IDS) that monitor network traffic for anomalous patterns can help detect and respond to potential attacks in real-time. By continuously monitoring the interconnects for suspicious activities, the system can proactively identify and mitigate security threats.

What other types of side channel attacks could be possible on multi-GPU systems, and how can they be detected and prevented?

In addition to congestion-based attacks, other types of side channel attacks could target multi-GPU systems, posing security risks. One potential attack vector is power consumption analysis, where adversaries analyze power usage patterns to infer information about the computations being performed on the GPUs. By monitoring power fluctuations during different operations, attackers could deduce sensitive data or cryptographic keys. To detect and prevent power consumption side channel attacks, implementing power monitoring mechanisms that track and analyze power usage in real-time can help identify abnormal patterns indicative of an attack. Moreover, introducing power management techniques that obfuscate power consumption profiles and reduce power differentials during computations can mitigate the effectiveness of such attacks.
Another type of side channel attack on multi-GPU systems could involve electromagnetic radiation emissions. Adversaries could exploit electromagnetic signals emitted by GPUs during processing to extract information about the data being processed. To detect and prevent electromagnetic side channel attacks, shielding the GPUs to minimize electromagnetic leakage and implementing electromagnetic interference (EMI) detection mechanisms can help safeguard against such threats. By conducting regular electromagnetic interference tests and deploying signal filtering techniques, the system can reduce the risk of data leakage through electromagnetic channels.

What are the potential implications of these attacks on the security and privacy of sensitive applications running on multi-GPU systems, such as those in the healthcare or financial sectors?

The implications of covert and side channel attacks on the security and privacy of sensitive applications running on multi-GPU systems, especially in sectors like healthcare and finance, are significant. In the healthcare sector, where patient data and medical records are highly confidential, these attacks could lead to unauthorized access to sensitive information. For instance, in a healthcare setting leveraging multi-GPU systems for medical imaging or patient data analysis, covert channel attacks could compromise patient privacy and confidentiality. Adversaries could intercept and decipher medical images or patient records, violating privacy regulations and potentially endangering patient well-being.
Similarly, in the financial sector, where secure processing of transactions and sensitive financial data is paramount, covert and side channel attacks on multi-GPU systems could have severe consequences. Attackers exploiting these vulnerabilities could gain access to financial transaction details, account information, or encryption keys, leading to financial fraud, data breaches, and compromised system integrity. The implications of such attacks in the financial sector extend beyond monetary losses to include reputational damage, legal repercussions, and regulatory non-compliance.
Overall, the potential implications of these attacks on sensitive applications in critical sectors like healthcare and finance underscore the importance of implementing robust security measures to safeguard data confidentiality, integrity, and availability on multi-GPU systems. Proactive security measures, regular vulnerability assessments, and continuous monitoring are essential to mitigate the risks posed by covert and side channel attacks in these high-stakes environments.