Enhancing Neural Network Robustness via Synthesis of Repair Programs
summary
A stochastic enumerative synthesizer named DefEnSyn generates short repair programs that change a few latent features via simple arithmetic assignments, learns promising neuron pairs by growing program length, and selects a diverse set to apply one at inference. Synthesis runs in about 1–2 hours on a few hundred images, defends against L_inf, L2, L0 and backdoor attacks, and yields robustness gains of +40% on average (up to +71%) while hurting clean accuracy by ≈1%, outperforming adversarial training, randomization, and prior repair methods.
AI-written summary · about
chapters (11)
- 0:00Neural network robustness intro
- 1:46Adversarial defense approaches
- 3:29Randomization repair limits
- 6:02Repair programs integration
- 8:50Repair programs synthesis
- 10:57Challenges in synthesis
- 12:01DefenseSynth algorithm
- 13:48Algorithm step-by-step
- 16:58Framework and evaluation
- 18:23Results and contributions
- 19:04Q&A on verification
related talks
18mActive Learning for Neurosymbolic Program Synthesis
01h 10mNeurosymbolic Program Synthesis: Bridging Perception and Reasoning in Real-World(…)
30mNeural Architecture Search using Property Guided Synthesis
21mArchitecture-Preserving Provable Repair of Deep Neural Networks
17mModeling Black-Box Components with Probabilistic Synthesis
38mProgram Synthesis: A Dream Realized?
29mSurvival of the Safest: Automating LLM Defense with Genetic Programming
16mFeedback-Driven Semi-supervised Synthesis of Program Transformations