Vulnerabilities in Deep Learning Language Models (DLLMs)
summary
Defines attack paths that actually change neuron weights during training: federated fine‑tuning spreads parameter updates over peers, enabling label‑flipping, backdoor triggers and programmatic model poisoning that survive long undetected; claims current large models use massive parameter counts like 175 billion parameters while some vendors report sourcing issues (uploader cites Hugging Face 99/100 models corrupted). Recommends detection, re‑training or full replacement of compromised models and strict partner vetting.
AI-written summary · about
chapters (7)
- 0:00DLLM security intro
- 6:47How LLMs learn language
- 15:03Neural networks & training
- 24:43Data poisoning risks
- 27:16Federated learning configs
- 32:55Supply chain risks
- 34:41Mitigation strategies
related talks
20mMachine Learning Attacks – A New Era of Security Epidemic
31mHacking with Words...
44mI Abuse Your LLM Live with 10 Security Vulnerabilities
02h 13mUnderstanding Security Threats Against Machine/Deep Learning Applications
28mSecurity of Machine Learning
23mSecurity and auditing tools in Large Language Models (LLM)
36mAdversarial Attacks on Deep Learning
21mTricking Neural Networks : Explore Adversarial Attacks