Found in Translation: A Generative Language Modeling Approach to Memory...
summary
FIT treats page access traces as a language and uses a generative encoder–decoder model (BERT with pretrained weights) to translate observed page sequences into object-level accesses, training on simulated requests from known distributions (CRIO display ads, medical DDX+, SIFT). It handles many-to-one page mappings and long-range correlations, achieves 71.7–99.9% accuracy across evaluated workloads, tolerates 1–10% page-tracking errors, runs against Nitro and SGX enclaves, and adds only milliseconds of added latency during online page monitoring.
AI-written summary · about
chapters (11)
- 0:00Memory access attack introduction
- 1:57Threat model and attack limits
- 5:06Correlations in memory accesses
- 7:33Language modeling for access translation
- 9:18FIT attack design and phases
- 11:04Language model architecture and training
- 12:31Evaluation on real-world apps
- 14:41Attack efficacy comparisons
- 16:05Practical error sensitivity
- 17:09Performance overhead impact
- 17:52Conclusion and future work
related talks
41mMerging Human Cyber Security Expertise with Generative AI
42mUnraveling AI terminology
22mCost-Efficient Large Language Model Serving for Multi-turn Conversations with...
43mBeyond the Hype: A Realistic Look at Large Language Models
37mBuilding Generative AI Applications
47mAI and Edge Computing: Efficient, Lean, Dependency‑Free Language Processing
01h 07mEmpowering Security Teams with Generative AI
15mEnabling Tensor Language Model to Assist in Generating High-Performance Tensor...