AGNES: Abstraction-guided Framework for Deep Neural Networks Security
summary
AGNES is a tool that converts models to ONNX, clusters hidden neurons and stimulates cluster representatives with an artificial brain stimulation optimizer to extract trigger patterns; it runs ~400 experiments on architectures (VGG, GoogleNet, etc.) with networks of 80k–3M parameters, finds more backdoors than ABS and Neural Cleanse, handles semi-transparent triggers, and reduces stimulation runtime by abstracting neurons via an approximate clustering / deep abstract pipeline.
AI-written summary · about
chapters (10)
- 0:00AGNES framework intro
- 1:03DNN security risks
- 3:11Motivation for AGNES
- 3:31Brain stimulation technique
- 4:13Clustering in AGNES
- 5:40Large network handling
- 6:43Experimental setup
- 8:28Results and backdoors
- 9:54Summary and future work
- 10:14Q&A on adversarial examples
related talks
19mAbstract Neural Networks
30mNeural Architecture Search using Property Guided Synthesis
19mSOTER: Guarding Black-box Inference for General Neural Networks at the Edge
09mAINETUS - AI for Safety-Critical Network Infrastructures
32mTowards new deep learning abstractions on top of exist frameworks
31mAbstract Neural Networks
36mBrain's Neural Network as a service
14mLightning Talk: A GNN Based Framework for Kubernetes Security Agents: Threat & Vulnerability...