Automatic Differentiation from Scratch: How PyTorch Computes Gradients in Physics-Informed Neural Networks

This collection of ML papers contains several findings with measurable security and privacy implications. Key points: (1) A detailed PyTorch autodiff trace for PINNs highlights how create_graph=True and multi-level differentiation build and retain complex computational graphs and adjoint values — increasing the attack surface for leakage of intermediate activations or sensitive inputs via gradients or retained graph state. (2) Federated Explainable AI (FedXAI) surveys explainability-in-flows and stresses explanation-centric privacy and robustness threats (non‑IID issues, leakage via explainers