Tianchen Zhang

Tianchen Zhang, Gururaj Saileshwar, David Lie

This paper demonstrates a new side-channel that enables an adversary to extract sensitive information about inference inputs in large language models (LLMs) based on the number of output tokens in the LLM response. We construct attacks using this side-channel in two common LLM tasks: recovering the target language in machine translation tasks and recovering the output class in classification tasks. In addition, due to the auto-regressive generation mechanism in LLMs, an adversary can recover the output token count reliably using a timing channel, even over the network against a popular closed-source commercial LLM. Our experiments show that an adversary can learn the output language in translation tasks with more than 75% precision across three different models (Tower, M2M100, MBart50). Using this side-channel, we also show the input class in text classification tasks can be leaked out with more than 70% precision from open-source LLMs like Llama-3.1, Llama-3.2, Gemma2, and production models like GPT-4o. Finally, we propose tokenizer-, system-, and prompt-based mitigations against the output token count side-channel.

Jiankun Wei, Abdulrahman Abdulrazzag, Tianchen Zhang et al.

Deployed large language models (LLMs) often rely on speculative decoding, a technique that generates and verifies multiple candidate tokens in parallel, to improve throughput and latency. In this work, we reveal a new side-channel whereby input-dependent patterns of correct and incorrect speculations can be inferred by monitoring per-iteration token counts or packet sizes.We demonstrate that an adversary observing these patterns can fingerprint user queries with >90% accuracy across four speculative-decoding schemes, REST (100\%), LADE (up to 92%), BiLD (up to 95%), and EAGLE (up to 77.6%) and leak confidential datastore contents used for prediction at rates exceeding 25 tokens/sec. We evaluate the side-channel attacks in both research prototypes as well as the production-grade vLLM serving framework. To defend against these, we propose and evaluate a suite of mitigations, including packet padding and iteration-wise token aggregation.

Guoming Zhang, Chen Yan, Xiaoyu Ji et al.

Speech recognition (SR) systems such as Siri or Google Now have become an increasingly popular human-computer interaction method, and have turned various systems into voice controllable systems(VCS). Prior work on attacking VCS shows that the hidden voice commands that are incomprehensible to people can control the systems. Hidden voice commands, though hidden, are nonetheless audible. In this work, we design a completely inaudible attack, DolphinAttack, that modulates voice commands on ultrasonic carriers (e.g., f > 20 kHz) to achieve inaudibility. By leveraging the nonlinearity of the microphone circuits, the modulated low frequency audio commands can be successfully demodulated, recovered, and more importantly interpreted by the speech recognition systems. We validate DolphinAttack on popular speech recognition systems, including Siri, Google Now, Samsung S Voice, Huawei HiVoice, Cortana and Alexa. By injecting a sequence of inaudible voice commands, we show a few proof-of-concept attacks, which include activating Siri to initiate a FaceTime call on iPhone, activating Google Now to switch the phone to the airplane mode, and even manipulating the navigation system in an Audi automobile. We propose hardware and software defense solutions. We validate that it is feasible to detect DolphinAttack by classifying the audios using supported vector machine (SVM), and suggest to re-design voice controllable systems to be resilient to inaudible voice command attacks.