Reversing Large Deep Learning AI Models
summary
Reversing extracts full layer graphs and numeric tensors from ONNX, HDF5 and PyTorch files so an attacker can download weights as numpy and read layer shapes, datatypes and sparsity; tools like Netron expose 22‑layer GoogLeNet internals, safe tensor formats reduce some serialization risks, and two repos (tensor, torture) provide tensor fuzzing and PyTorch poisoning. Small weight perturbations in dense matrices (often sparsity 0%) reliably break vision models; control‑theory analyses (root‑locus, Bode) identify the most susceptible layers.
AI-written summary · about
chapters (8)
- 0:00Introduction to AI model reversing
- 4:44AI model threat modeling foundations
- 8:41Deep learning model formats overview
- 13:00Reversing GoogleNet architecture
- 24:32AI as feedback control systems
- 30:30Demo of weight manipulation attack
- 33:01AI security core principles summary
- 33:21Q&A session
related talks
01h 05mDeep Learning with PyTorch
27mDemystifying Deep Neural Networks
25mTricking Neural Networks : Explore Adversarial Attacks
52mSleeping With One AI Open: An Introduction to Attacks Against...
36mVulnerabilities in Deep Learning Language Models (DLLMs)
32mDnD: Decompiling Deep Neural Network Compiled Binary
29mBarraCUDA: GPUs Do Leak DNN Weights
35mDeep Learning Model Interpretability for Computer Vision based Models