Official repository for CVPR'23 paper: Detecting Backdoors in Pre-trained Encoders

BackdoorMBTI is an open source project expanding the unimodal backdoor learning to a multimodal context. We hope that BackdoorMBTI can facilitate the analysis and development of backdoor defense methods within a multimodal context.

This repository is the official implementation of the paper "ASSET: Robust Backdoor Data Detection Across a Multiplicity of Deep Learning Paradigms." ASSET achieves state-of-the-art reliability in detecting poisoned samples in end-to-end supervised learning/ self-supervised learning/ transfer learning.

[CCS'24] Official Implementation of "Fisher Information guided Purification against Backdoor Attacks"

[TPAMI 2025] Dynamic Attention Analysis for Backdoor Detection in Text-to-Image Diffusion Models

Unofficial implementation of the detector cleanse in the paper "BadDet: Backdoor Attacks on Object Detection."

Educational research demonstrating weight manipulation attacks in SafeTensors models. Proves format validation alone is insufficient for AI model security.

Rubber Ducky is an awareness initiative focused on educating individuals and organizations about potential cybersecurity threats that can be delivered through USB devices, particularly targeting issues like ransomware, backdoors, and keyloggers. Leveraging the concept of Bad USB and the functionality of Rubber Ducky USBs.

Solution for the Third Challenge of the Main Phase in the Rayan International AI Contest: Backdoored Model Detection.

This repository contains the code for the paper titled "GenLLMGuard: Detecting Backdoors in LLMs for Open-Ended Text Generation Through Trigger Inversion".

DEB: A Two-Stage Backdoor Defense for Split Neural Network-Based Vertical Federated Learning

General Test-Time Backdoor Detection in Split Neural Network-Based Vertical Federated Learning

SaveX is a Python library crafted to bolster the security of file uploads within web applications. With an emphasis on mitigating prevalent vulnerabilities associated with file handling, SaveX furnishes sturdy sanitization and validation capabilities to ensure that uploaded files are safe for utilization within your application.

Inspired by dynamic taint tracking, PoisonSpot uses fine-grained training provenance tracker that: (1) tags & traces the impact of every single training sample on model updates, (2) probabilistically scores suspect samples based on their lineage of impact on model weights, and (3) separates the clean from the poisonous before retraining a model.