Adversarial Attacks Undermine Explanation Robustness of Rationalization Models
Existing rationalization models are vulnerable to adversarial attacks that can significantly change the selected rationales while maintaining model predictions, undermining the credibility of these models.
Backdoor Attacks Against Speaker Identification Using Emotional Prosody
Emotional prosody can be effectively used as a backdoor trigger to compromise the integrity of speaker identification deep neural networks.
Evaluating the Robustness of Phishing Webpage Detection Models against Adversarial Attacks Generated by PhishOracle
PhishOracle, a tool that generates adversarial phishing webpages by embedding diverse content-based and visual-based phishing features into legitimate webpages, can be used to evaluate the robustness of existing phishing webpage detection models.
Interpreting Foundation Models as Compressed Representations of Training Data: Implications for Copyright Law
The training process of foundation models can be interpreted as a form of data compression, where the model's weights represent a compressed version of the training data. This perspective has significant implications for understanding the copyright status of the model weights and the outputs generated by the model.
Defending Against Clean-Label Backdoor Attacks in Cybersecurity Machine Learning Models
A novel defense mechanism that leverages density-based clustering and iterative scoring to effectively mitigate clean-label backdoor attacks on machine learning models used in cybersecurity applications, without requiring access to clean training data or knowledge of the victim model architecture.
Evaluating the Impact of Censorship and Domain Adaptation on the Detection of Machine-Generated Tweets
Censorship and domain adaptation significantly undermine the effectiveness of automated detection methods in identifying machine-generated tweets.
Quantitative Attack-Defense Tree Synthesis, Analysis and Verification
Quantitative analysis of attack-defense trees can distinguish likely from unlikely vulnerabilities by utilizing information such as probabilities, costs, and timing. This paper presents a tool, QuADTool, that allows for easy synthesis and analysis of attack-defense tree models, including support for probabilities, costs, and time. The tool also provides interfaces to existing model checkers and analysis tools.
인공지능 기반 온라인 선거 개입의 위험한 활용 양상 분석
제너레이티브 인공지능(GenAI)과 대규모 언어 모델(LLM)은 온라인 선거 개입을 위한 심각한 위험을 초래할 수 있다.
Generative AI Poses Significant Risks for Online Election Interference: Exploring Nefarious Applications and Mitigation Strategies
Generative Artificial Intelligence (GenAI) and Large Language Models (LLMs) present significant risks for online election interference, enabling sophisticated forms of manipulation and disinformation that can disrupt democratic processes.
Probabilistic Approach for Robust Watermarking of Neural Networks
A novel probabilistic approach for enhancing the robustness of trigger set-based watermarking techniques against model stealing attacks.
© 2024 by Linnk AI