Trusted-AI/adversarial-robustness-toolbox

Features

• Adversarial Machine Learning - Provides a unified library for testing and hardening machine learning models against evasion, poisoning, and extraction attacks.
• Adversarial Example Generations - Crafts evasion attacks by computing input gradients through the model and perturbing along the direction that maximizes loss.
• Extraction Model Evaluation - Tests model resilience to attackers reconstructing a functional copy by querying it.
• Model Extraction Attack Evaluations - Tests model resilience to adversaries reconstructing a functional copy by querying it.
• Model Functionality Extraction - Replicates target model behavior by querying it and training a substitute on the responses.
• Adversarial Robustness Testing - Generates adversarial examples using PGD, FGSM, and DeepFool to test model prediction robustness.
• Adversarial Robustness Libraries - Provides a unified library for testing and hardening models against adversarial attacks across multiple ML frameworks.
• Adversarial Example Augmentation - Generates adversarial examples during training to harden models against crafted input perturbations.
• Adversarial Robustness Training - Trains models with adversarial examples to improve classification accuracy on subtly perturbed inputs.
• Adversarial Robustness Trainings - Augments training data with adversarial examples to improve a model's resistance to crafted inputs during inference.
• Randomized Smoothing Certifications - Applies noise injection and statistical bounds to produce provable robustness guarantees around a model's predictions.
• Multi-Framework Model Collections - Works with models built in TensorFlow, PyTorch, scikit-learn, and other popular frameworks without requiring code changes.
• Robustness Certification Tools - Provides formal verification and randomized smoothing to certify model predictions within defined perturbation bounds.
• Robustness Certifications - Provides formal guarantees that a model's predictions remain stable within a defined perturbation radius around any input.
• Inference Attack Evaluations - Tests model resilience to adversaries inferring sensitive attributes or training data membership.
• Adversarial Attacks - Provides a framework for generating evasion, poisoning, and extraction attacks to stress-test model robustness.
• Attack-Defense Pipelines - Provides composable pipelines that chain attack generation, defense application, and robustness evaluation on any wrapped estimator.
• Adversarial Input Detection - Analyzes inputs and activations to flag samples crafted to deceive the model.
• Adversarial Threat Defenses - Applies preprocessing, postprocessing, and detection techniques to harden models against multiple attack types.
• Evasion Attack Evaluations - Tests model resilience to adversarial inputs designed to cause misclassification during inference.
• Evasion Attack Generators - Generates evasion attacks against classification models, optimizing strength on correctly classified samples.
• Model Extraction Prevention - Simulates model extraction attacks and provides countermeasures to protect intellectual property.
• Adversarial Input Sanitization - Applies transformations like compression and smoothing to strip adversarial perturbations before inference.
• Inference Attack Prevention - Identifies and blocks attempts to infer sensitive information like membership or attributes from a model.
• Model Inference Attacks - Extracts private training data attributes or membership status by analyzing a model's predictions.
• ML Framework Abstractions - Applies adversarial robustness techniques across models built in TensorFlow, PyTorch, scikit-learn, and other frameworks without code changes.
• Framework Abstraction Layers - Wraps models from different frameworks behind a common estimator interface so attacks and defenses work without framework-specific code.
• Classification Models - Wraps classification models from Keras, PyTorch, and scikit-learn under a unified estimator interface.
• Training Data Reconstruction Attacks - Recovers missing rows of a training dataset by exploiting a model's learned representations.
• Object Detection - Ships object detection estimators for architectures like Faster-RCNN and YOLO under a unified interface.
• Regression Models - Wraps regression estimators from Keras, PyTorch, and scikit-learn under a unified interface.
• Speech to Text Transcription - Runs automatic speech recognition on audio using PyTorch or TensorFlow estimators.
• Video Object Tracking - Runs object tracking on video frames using a PyTorch estimator for the GOTURN architecture.
• Privacy Attack Evaluations - Assesses a model's vulnerability to inference, membership, and reconstruction attacks that leak sensitive training data.
• Model Extractions - Replicates a target model's behavior by querying it and training a local surrogate on the collected predictions.
• Poisoning Attack Implementations - Injects malicious data into the training set to compromise a model's behavior, including backdoors and clean-label triggers.
• Spectral Signature Detections - Detects and mitigates malicious training data using spectral signatures and other poisoning defense techniques.
• Poisoning Attack Evaluations - Tests a model's resilience to corrupted training data that degrades its performance or introduces backdoors.
• Poisoning Detections - Identifies and mitigates malicious training data injected to corrupt model behavior or introduce backdoors.
• Secure - Applies transformations to raw predictions to obscure confidence scores or reduce information leakage to attackers.
• Poisoning Detections - Identifies malicious training samples by analyzing covariance structure and outlier scores in the feature space.
• Privacy Audits - Probes a model's outputs to infer membership or attribute information, quantifying leakage risk.
• Model Extraction Protections - Simulates and counters model extraction attacks to protect intellectual property in deployed ML systems.
• Preprocessing Sanitizations - Transforms raw inputs through compression, smoothing, or encoding to strip adversarial perturbations before inference.
• AI Security - Library for testing ML model robustness against adversarial attacks.
• AI Security and Red Teaming - Python library for adversarial machine learning.
• Frameworks and Benchmarks - General-purpose library for adversarial and backdoor robustness testing.
• Sample Filtering - Provides activation clustering to detect backdoors in deep networks.
• Research Toolkits - Library for adversarial robustness and backdoor defense evaluation.
• Guardrails and AI Safety - Listed in the “Guardrails and AI Safety” section of the The Incredible Pytorch awesome list.
• Privacy and Safety - Library for defending and evaluating ML model robustness.
• Security And Privacy - Security library for machine learning models.

Star history