
SynGhost
Syntactic Ghost: An Imperceptible General-purpose Backdoor Attacks on Pre-trained Language Models

Syntactic Ghost: An Imperceptible General-purpose Backdoor Attacks on Pre-trained Language Models

Embed multiple secret messages in LLM chat token choices using arithmetic/Discop steganographic coders, with bit-exact decoding and steganalysis…

Prohibited Items Segmentation via Occlusion-aware Bilayer Modeling (ICME 2025)

Research code for training adapters that make LLM agent backdoors survive benign fine-tuning, with trigger-rate scoring, SWE-bench evaluation, and…

A novel adversarial attack on LLM based on the Exponentiated Gradient Descent technique.

CLPSTNet: A Progressive Multi-Scale Convolutional Steganography Model Integrating Curriculum Learning

Benchmark harness measuring where prompt injection defenses fire in tool-using LLM agent pipelines, tracking canary tokens across exposed, persisted,…

Research implementation of the Black-Hole Attack that poisons vector database indexes via centrality-driven hubness, with FAISS pipelines and…

Reference implementation of privacy-preserving face recognition using replicated secret sharing in MP-SPDZ, with quantized GhostFaceNet inference,…

Research code, benchmark and defense for target rebinding attacks that trick LLM safety guards into grading forged decoy content, plus segment-max…

Research implementation of Hop-Decayed Influence (HDI) and the 3S attack framework, exposing structural auxiliary indexing vulnerabilities in…

Experiments for control-token chain-of-thought suppression and parser-leniency attacks on tool-using LLM agents

How Much Can Tool Calls Reveal About You? Privacy Leakage from Tool-Call Traces

Research code implementing Dynamical Low-Rank Nash Equilibrium computation for stochastic ICS security games between advanced persistent threats and…

Adversarial image attacks on vision-language web agents, from visual grounding to browser execution

Hide valuable information where it's least expected

Public version of the Entropica repo

Research code for adding controllable safety experts to LLM layers via two-stage upcycling training, with inference-time safety temperature routing…