Data Contamination and Leakage Detection
Research on identifying, preventing, and mitigating data contamination and leakage in training and evaluation of large language models
Data Contamination and Leakage Detection
Research on Large Language Models in Data Contamination and Leakage Detection
Exposing the Blind Spot in Multimodal LLMs
A Framework to Detect Data Contamination Across Text & Images
Finding Hidden Memories in Large Language Models
Automated Detection of Privacy Vulnerabilities at Scale
Quantifying Data Extraction Risks in LLMs
A sequence-level probability approach to measuring training data leakage
Security Risks in Code Language Models
Investigating Data Extraction Vulnerabilities Before and After Fine-tuning
Uncovering Dataset Contamination in LLMs
A new metric for measuring training data leakage into evaluation sets
The Hidden Danger in AI Evaluation
How LLMs judging other LLMs creates security vulnerabilities
Securing LLMs from Toxic Training Data
A Data Attribution Approach to Finding & Filtering Unsafe Content
Privacy at Risk: Stealing Personal Data from LLMs
New technique extracts personally identifiable information from language models
Rethinking Data Poisoning in LLMs
From Security Vulnerabilities to Development Opportunities
Safeguarding AI Evaluation Integrity
Detecting benchmark contamination with innovative watermarking techniques
Evolving LLM Benchmarks
From Static to Dynamic Evaluation: Combating Data Contamination
Poison Pills in LLMs: Hidden Vulnerabilities
How targeted data poisoning compromises AI security
Detecting LLM Training Data Exposure
New Attack Method Requires Only Generated Outputs
Defending Against Dead Code Poisoning
Novel detection techniques to secure code generation models
The Hidden Dangers of Private LLM Evaluations
Security risks in closed-door model assessments
Unveiling the Black Box of LLM Training Data
A novel approach to detect data imprints in proprietary models
The IP Protection Dilemma in LLM Fine-Tuning
Balancing utility and intellectual property protection for hardware design
The Illusion of LLM Benchmark Success
Revealing the failures of contamination mitigation strategies
The LLM Memorization Challenge
How language models can complete texts they weren't explicitly trained on
Defending Against Data Poisoning
Understanding threats to deep learning security
Security Risks in Code-Generating LLMs
Uncovering sensitive information disclosure vulnerabilities
Password Vulnerabilities in Fine-tuned LLMs
How sensitive data can leak through model parameters
Rethinking Data Markets for LLMs
Game theory reveals flaws in current data valuation methods
