Research software for neural network verification, model analysis, and ONNX workflows.
Published tools with accompanying write-ups.
Key-based model usage control with adaptation restricted to an intrinsic subset of parameters, avoiding full re-keying within that update procedure.
Retraining-free model behaviour modulation through logits redistribution, including utility and focus modulation.
Recoding data in designated-model low-sensitivity directions to preserve its task utility and study reduced transfer to other models.
Model usage control by extracting selected influential weights as a key for restoring the model's original capability.
Token obfuscation to reduce gradient-based text reconstruction while retaining useful training signals in evaluated language-model settings.
Gradient inversion in language-model training, combining dropout-mask estimation with discrete token-sequence optimisation.
Character-level perturbation specifications and controlled text augmentation for studying language-model robustness.
A shadow-query representation for reducing document-reconstruction risks from stored embeddings while retaining retrieval utility.
Efficiently constructing tight over-approximations for activation function hulls, with reported average 400x faster construction on benchmarked Sigmoid, Tanh, MaxPool, and related cases.
A fast and precise approach to over-approximating the convex hull of the ReLU function, reporting up to 50% fewer constraints at comparable or better precision.
Supporting infrastructure for ONNX and verification workflows.
A unified Python library to approximate activation function hull with convex polytopes. Supports ReLU, LeakyReLU, ELU, Sigmoid, Tanh, and MaxPool.
View on GitHubA tool to infer missing tensor shapes in ONNX models for inspection and downstream tooling.
View on GitHubA tool to optimize and simplify your ONNX models by removing redundant operations.
View on GitHubA tool for converting ONNX models to PyTorch models (`.pth` for parameters, `.py` for structure).
View on GitHubA tool to convert VNN-LIB files (`.vnnlib`) to PyTorch tensors (`.pth` files) for efficient neural network verification.
View on GitHub