Michael Backes
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Michael Backes has a total of 58 co-authors. Here are the top co-authors by number of shared publications:
- Yang Zhang
CISPA Helmholtz Center for Information Security, Saarbrücken, Germany
14 shared publications - Xinyue Shen
University of Waterloo, Waterloo, Canada and CISPA Helmholtz Center for Information Security, Saarbrücken, Germany
6 shared publications - Yun Shen
Flexera, Bracknell, U.K.
5 shared publications - Yiting Qu
CISPA Helmholtz Center for Information Security, Saarbrücken, Germany
3 shared publications - Yixin Wu
CISPA Helmholtz Center for Information Security
3 shared publications - Zeyuan Chen
CISPA Helmholtz Center for Information Security, Saarbrücken, Germany
3 shared publications
Latest Publications
- Open Access
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 287-301
'Shut Up and Let Me Enjoy My Otome': Understanding and Measuring the Toxicity in Otome Game Communities
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 287-301
Open Access'Shut Up and Let Me Enjoy My Otome': Understanding and Measuring the Toxicity in Otome Game Communities
- Open Access
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 4432-4446
Real Money, Fake Models: Deceptive Model Claims in Shadow APIs
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 4432-4446
Open AccessReal Money, Fake Models: Deceptive Model Claims in Shadow APIs
- Open Access
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 4417-4431
PersonaHate: A Scalable Persona-Based Data Synthesis Pipeline for Hate Speech Analysis
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 4417-4431
Open AccessPersonaHate: A Scalable Persona-Based Data Synthesis Pipeline for Hate Speech Analysis
- 15 September 2026
- Published by Springer Nature
- p. 114-131
MultiMem: Measuring and Mitigating Memorization in Multi-modal Contrastive Learning
- 15 September 2026
- Published by Springer Nature
- p. 114-131
MultiMem: Measuring and Mitigating Memorization in Multi-modal Contrastive Learning
- 10 September 2026
- Published by Springer Nature
- p. 547-565
GEO-Detective: Unveiling Location Privacy Risks in Images with LLM Agents
- 10 September 2026
- Published by Springer Nature
- p. 547-565
GEO-Detective: Unveiling Location Privacy Risks in Images with LLM Agents
- 1 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. 21 (15566013) , 5530-5543
BadBone: Backdoor Attacks Against Backbone Models in Visual Prompt Learning
Prompt learning is a new machine learning paradigm that has attracted ample attention due to its simplicity and proven efficacy. Despite its growing adoption, the security vulnerabilities associated with this paradigm remain underexplored. In this work, we take the first step to propose BadBone, a stealthy and adaptive backdoor attack against prompt learning using bi-level optimization. Instead of backdooring the prompt learning process, we aim to compromise a backbone model such that only target downstream tasks employing prompt learning inherit the backdoor vulnerability. Extensive experiments on three different models and three datasets from various domains show that our targeted/untargeted backdoored models achieve high attack performance while maintaining utility on both pretraining and downstream tasks. Moreover, we evaluate our approach against six state-of-the-art model-level defenses, including Neural Cleanse, ABS, MNTD, NAD, CLP, and D-BR. The results demonstrate that these defenses are largely ineffective against our backdoored models and thus leave the effective defense as an important direction for future work. Our code is available at https://github.com/TrustAIRLab/BadBone.
- 1 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. 21 (15566013) , 5530-5543
BadBone: Backdoor Attacks Against Backbone Models in Visual Prompt Learning
Prompt learning is a new machine learning paradigm that has attracted ample attention due to its simplicity and proven efficacy. Despite its growing adoption, the security vulnerabilities associated with this paradigm remain underexplored. In this work, we take the first step to propose BadBone, a stealthy and adaptive backdoor attack against prompt learning using bi-level optimization. Instead of backdooring the prompt learning process, we aim to compromise a backbone model such that only target downstream tasks employing prompt learning inherit the backdoor vulnerability. Extensive experiments on three different models and three datasets from various domains show that our targeted/untargeted backdoored models achieve high attack performance while maintaining utility on both pretraining and downstream tasks. Moreover, we evaluate our approach against six state-of-the-art model-level defenses, including Neural Cleanse, ABS, MNTD, NAD, CLP, and D-BR. The results demonstrate that these defenses are largely ineffective against our backdoored models and thus leave the effective defense as an important direction for future work. Our code is available at https://github.com/TrustAIRLab/BadBone.
- Top Cited Papers
- 9 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. 21 (15566013) , 3436-3449
Robustness Over Time: Understanding Adversarial Examples’ Effectiveness on Longitudinal Versions of Large Language Models
Large Language Models (LLMs) undergo continuous updates to improve user experience. However, prior research on the security and safety implications of LLMs has primarily focused on their specific versions, overlooking the impact of successive LLM updates. This prompts the need for a holistic understanding of the risks in these different versions of LLMs. To fill this gap, in this paper, we conduct a longitudinal study to examine the adversarial robustness – specifically misclassification, jailbreak, and hallucination – of three prominent LLM families: GPT, Llama, and Qwen. Our study reveals that LLM updates do not consistently improve adversarial robustness as expected. For instance, a later version of GPT-3.5 degrades regarding misclassification and hallucination despite its improved resilience against jailbreaks. GPT-4 and GPT-4o demonstrate (incrementally) higher robustness overall. Larger Llama and Qwen models do not uniformly exhibit improved robustness across all three aspects studied. In addition, larger model sizes do not necessarily yield improved robustness. Minor updates lacking substantial robustness improvements can exacerbate existing issues rather than resolve them. We hope our study can offer valuable insights into navigating model updates and informed decisions in model development and usage.
- 9 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. 21 (15566013) , 3436-3449
Top Cited PapersRobustness Over Time: Understanding Adversarial Examples’ Effectiveness on Longitudinal Versions of Large Language Models
Large Language Models (LLMs) undergo continuous updates to improve user experience. However, prior research on the security and safety implications of LLMs has primarily focused on their specific versions, overlooking the impact of successive LLM updates. This prompts the need for a holistic understanding of the risks in these different versions of LLMs. To fill this gap, in this paper, we conduct a longitudinal study to examine the adversarial robustness – specifically misclassification, jailbreak, and hallucination – of three prominent LLM families: GPT, Llama, and Qwen. Our study reveals that LLM updates do not consistently improve adversarial robustness as expected. For instance, a later version of GPT-3.5 degrades regarding misclassification and hallucination despite its improved resilience against jailbreaks. GPT-4 and GPT-4o demonstrate (incrementally) higher robustness overall. Larger Llama and Qwen models do not uniformly exhibit improved robustness across all three aspects studied. In addition, larger model sizes do not necessarily yield improved robustness. Minor updates lacking substantial robustness improvements can exacerbate existing issues rather than resolve them. We hope our study can offer valuable insights into navigating model updates and informed decisions in model development and usage.
- Top Cited PapersOpen Access
- 19 November 2025
- Published by Association for Computing Machinery (ACM)
- p. 3221-3235
UnsafeBench: Benchmarking Image Safety Classifiers on Real-World and AI-Generated Images
- 19 November 2025
- Published by Association for Computing Machinery (ACM)
- p. 3221-3235
Top Cited PapersOpen AccessUnsafeBench: Benchmarking Image Safety Classifiers on Real-World and AI-Generated Images
- Open Access
- 19 November 2025
- Published by Association for Computing Machinery (ACM)
- p. 3102-3116
DivTrackee versus DynTracker: Promoting Diversity in Anti-Facial Recognition against Dynamic FR Strategy
- 19 November 2025
- Published by Association for Computing Machinery (ACM)
- p. 3102-3116
Open AccessDivTrackee versus DynTracker: Promoting Diversity in Anti-Facial Recognition against Dynamic FR Strategy
- 6 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Dependable and Secure Computing
- Vol. 23 (1) , 1514-1527
Revealing the Risk of Hyper-Parameter Leakage in Deep Reinforcement Learning Models
Deep reinforcement learning (DRL) has been implemented across various critical applications, including smart grids, trac management systems, and autonomous vehicles. To safeguard intellectual property and mitigate security vulnerabilities, access to DRL models is typically restricted to a black-box format. is means specic details like the structure of the policy network and optimization processes are not openly available to users. It is crucial to determine if the hyper-parameters can be inferred from observable states and actions within these models, presenting two primary challenges: 1) limited data available from the black-box model and 2) the intertwined eects of hyperparameters on the model's behavior. Since DRL models exhibit varying behaviors in identical tasks depending on their hyper-parameter congurations, we introduce a novel hyper-parameter inference attack against DRL, named HyperInfer, which allows adversaries to deduce the settings of a black-box DRL model. In order to fully assess the risk of model hyper-parameter leakage, we design two novel state generation methods that provoke divergent responses from DRL models. We also develop an inference framework to elucidate the relationship between model behavior and hyper-parameter settings. rough comprehensive experiments involving multiple DRL models and environments, we demonstrate that model behaviors can indeed reveal hyper-parameter settings, with inference accuracy surpassing 90% in scenarios such as PPO with CartPole. We also discuss keyndings relevant to practical applications and explore how knowledge of hyperparameters can facilitate more sophisticated attacks. Lastly, we propose potential defensive strategies to minimize the risk of hyper-parameter leakage in DRL models.
- 6 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Dependable and Secure Computing
- Vol. 23 (1) , 1514-1527
Revealing the Risk of Hyper-Parameter Leakage in Deep Reinforcement Learning Models
Deep reinforcement learning (DRL) has been implemented across various critical applications, including smart grids, trac management systems, and autonomous vehicles. To safeguard intellectual property and mitigate security vulnerabilities, access to DRL models is typically restricted to a black-box format. is means specic details like the structure of the policy network and optimization processes are not openly available to users. It is crucial to determine if the hyper-parameters can be inferred from observable states and actions within these models, presenting two primary challenges: 1) limited data available from the black-box model and 2) the intertwined eects of hyperparameters on the model's behavior. Since DRL models exhibit varying behaviors in identical tasks depending on their hyper-parameter congurations, we introduce a novel hyper-parameter inference attack against DRL, named HyperInfer, which allows adversaries to deduce the settings of a black-box DRL model. In order to fully assess the risk of model hyper-parameter leakage, we design two novel state generation methods that provoke divergent responses from DRL models. We also develop an inference framework to elucidate the relationship between model behavior and hyper-parameter settings. rough comprehensive experiments involving multiple DRL models and environments, we demonstrate that model behaviors can indeed reveal hyper-parameter settings, with inference accuracy surpassing 90% in scenarios such as PPO with CartPole. We also discuss keyndings relevant to practical applications and explore how knowledge of hyperparameters can facilitate more sophisticated attacks. Lastly, we propose potential defensive strategies to minimize the risk of hyper-parameter leakage in DRL models.