SoK: Efficiency Robustness of Dynamic Deep Learning Systems
summary
Paper defines a taxonomy of efficiency attacks on adaptive models across D1: skipping mechanisms, D2: dynamic inference iterations, and D3: variable output production, and demonstrates concrete hits: multiexit skipping attacks cut successful inferences from ~10,000 to ~3,500 on a Galaxy S9 Plus and cause up to 5× latency; ODEN solver perturbations inflate integration steps and raise cost by ~50%; autoregressive attacks blow output from 10 to 62 tokens (JPEG reduces to 11 but with poor BLEU); UAPs slow video pipelines from 40 to 16 FPS. Defenses like JPEG, smoothing, and an SVM detector give only partial or fragile protection.
AI-written summary · about
chapters (17)
- 0:00Dynamic deep learning systems
- 1:16Efficiency attacks and risks
- 2:22Dynamic behaviors in DL
- 3:47Attacker goals and capabilities
- 5:11Taxonomy of efficiency attacks
- 5:31Skipping attacks in D1
- 8:05Impact of skipping attacks
- 8:26Step size attacks on ODE networks
- 9:31Spiking neural network vulnerabilities
- 10:58Sparsity attacks on accelerators
- 12:45Dynamic inference in D2
- 14:55Impact of D2 attacks
- 15:59Attacks on object detection in D3
- 17:45Impact of D3 attacks
- 18:06Defenses against efficiency attacks
- 19:54Mitigation effectiveness results
- 20:37Key takeaways and future work
related talks
36mAdversarial Attacks on Deep Learning
02h 13mUnderstanding Security Threats Against Machine/Deep Learning Applications
48mStay Ahead of Adversarial AI in OT/ICS Environments – Mitigating CWE-1039
21mAdversarial Attacks
36mAdversarial Machine Learning: How Secure Is ML?
46mLessons Learned from Evaluating the Robustness of Defenses to Adversarial Examples
40mEvasion, Poisoning, Extraction, and Inference: Tools to Defend and Evaluate
25mTricking Neural Networks : Explore Adversarial Attacks