AIware 2024
Mon 15 - Tue 16 July 2024 Porto de Galinhas, Brazil, Brazil
co-located with FSE 2024

Large language models (LLMs) have revolutionized software development practices, yet concerns about their safety have arisen, particularly regarding hidden backdoors, \textit{aka} trojans. Backdoor attacks involve the insertion of triggers into training data, allowing attackers to manipulate the behavior of the model maliciously. In this paper, we focus on analyzing the model parameters to detect potential backdoor signals in code models. Specifically, we examine attention weights and biases, and context embeddings of the clean and poisoned CodeBERT and CodeT5 models. Our results suggest noticeable patterns in context embeddings of poisoned samples for both the poisoned models; however, attention weights and biases do not show any significant differences. This work contributes to ongoing efforts in white-box detection of backdoor signals in LLMs of code through the analysis of parameters and embeddings.

Mon 15 Jul

Displayed time zone: Brasilia, Distrito Federal, Brazil change

16:00 - 18:00
Security and Safety + Round Table + Day1 ClosingMain Track / Late Breaking Arxiv Track at Mandacaru
Chair(s): Thomas Zimmermann Microsoft Research, Ahmed E. Hassan Queen’s University
16:00
5m
Paper
An AI System Evaluation Framework for Advancing AI Safety: Terminology, Taxonomy, Lifecycle Mapping
Main Track
Boming Xia CSIRO's Data61 & University of New South Wales, Qinghua Lu Data61, CSIRO, Liming Zhu CSIRO’s Data61, Zhenchang Xing CSIRO's Data61
DOI
16:05
5m
Paper
Measuring Impacts of Poisoning on Model Parameters and Embeddings for Large Language Models of Code
Main Track
Aftab Hussain University of Houston, Md Rafiqul Islam Rabin University of Houston, Amin Alipour University of Houston
DOI
16:10
10m
Paper
A Case Study of LLM for Automated Vulnerability Repair: Assessing Impact of Reasoning and Patch Validation Feedback
Main Track
Ummay Kulsum North Carolina State University, Haotian Zhu Singapore Management University, Bowen Xu North Carolina State University, Marcelo d'Amorim North Carolina State University
DOI
16:20
5m
Paper
Trojans in Large Language Models of Code: A Critical Review through a Trigger-Based Taxonomy
Late Breaking Arxiv Track
Aftab Hussain University of Houston, Md Rafiqul Islam Rabin University of Houston, Toufique Ahmed University of California at Davis, Bowen Xu North Carolina State University, Premkumar Devanbu UC Davis, Amin Alipour University of Houston
Pre-print
16:25
25m
Live Q&A
Session Q&A and topic discussions
Main Track

16:50
60m
Panel
Round Table
Main Track

17:50
10m
Day closing
Day 1 summary and closing
Main Track