Fun to work with @JaechulRoh on this project.
Context:
There is a growing interest in Audio Language Models from Industry.
Question:
Can we fool these models via adversarial audio samples despite degradation from codec compression?
TLDR: Yes ;)
Excited for what’s next!
New preprint: Codec-Robust Attacks on Audio LLMs
#CodecAttack
Lossy codecs (Opus, MP3, AAC) have been treated as a defense against adversarial audio. We show they're actually an attack surface.
New preprint: Codec-Robust Attacks on Audio LLMs
#CodecAttack
Lossy codecs (Opus, MP3, AAC) have been treated as a defense against adversarial audio. We show they're actually an attack surface.
Saw a student in the plane writing the prior art of a paper with the assistance of GPT.
Made me wonder if it still makes sense to have SoK papers in the era of GPT and co.
Thoughts?
Heading to Singapore for AAAI
We will be presenting our work « CP-Freezer » which is a latency attack targeting Vehicular Cooperative Perception.
Paper: https://t.co/fV1VHB0Fn4
Hope to see you new faces there :)
7/n
Our intent is to propose an evaluation framework to evaluate the robustness of the attack based on the characteristic of the AI system under attack.
We hope our past experiences will help the community to raise the bar, at least, for latency attacks!
Our #Qualcomm AI Security Research team will be presenting our latest (published) work at #NeurIPS2025
Paper: https://t.co/uEhDu0qeCe
Poster session: Wed 3 Dec 4:30 p.m. PST — 7:30 p.m. PST
Qualcomm booth: Thur 4 Dec 9:45 am PST
Feel free to reach out!
More context below :)
6/n
For instance, this targeted system could have:
- a quantized version of the model
- a model of different size
- a model with a different architecture
- a system with some defenses
- different types of hardware and compute power
Heading to Neurips with our Qualcomm AI Security Research team.
We will be presenting our paper on evaluating adversarial latency attacks.
Happy to connect if you want to collaborate or/and are looking for internships (2027) around topics related to AI Security / Safety :)