Making Acoustic Side-Channel Attacks on Noisy Keyboards Viable with LLM-Assisted...
summary
Vision transformers replace CNNs and reach or beat the CoAtNet baseline on keystroke spectrogram classification, while additive Gaussian noise causes the prior QuatNet to fall from 99% accuracy to much lower numbers with just ~10% noise. A pipeline then runs classification outputs through an LLM for error correction; GPT-4o substantially improves sentence recovery and a fine-tuned Llama 3B with LoRA delivers ~90% of GPT-4o’s correction performance at dramatically smaller model size, making the attack practical.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 0:35Acoustic side-channel attacks matter
- 1:40Passphrases vulnerable to audio
- 2:46Noise challenges acoustic attacks
- 3:53Noise impact on model accuracy
- 4:35Two-stage pipeline for keystroke recovery
- 6:21Vision preprocessing and spectrograms
- 7:04Fine-tuning CoAtNet and vision transformers
- 8:10Robustness against synthetic noise
- 8:54LLM fixes typos from spectrograms
- 9:41LLM performance evaluation
- 10:44LLM resource constraints
- 11:28Fine-tuning Llama 3B
- 12:13Gaps in attack datasets
- 12:56Call for realistic datasets
- 13:18Conclusion and security implications
- 14:03Mitigating acoustic side-channel risks
related talks
45mThe Sound of Your Secrets: Teaching Your Model to Spy So You Can Learn to Defend
26mModern NLP for Proactive Harmful Content Moderation
49mChatGPT: A New Generation of Dynamic Machine Based Attacks?
48mKeynote – Sameer Singh
01h 04mRobust Generalization in the Era of LLMs: Jailbreaking Attacks and Defenses
46mHow Large Language Models Are Reshaping the Cybersecurity Landscape
25mTricking Neural Networks : Explore Adversarial Attacks
28mIndirect Prompt Injection Into LLMs Using Images and Sounds