Torsten Krauß

Phillip Rieger, Torsten Krauß, Markus Miettinen et al.

Federated Learning (FL) is a promising approach enabling multiple clients to train Deep Neural Networks (DNNs) collaboratively without sharing their local training data. However, FL is susceptible to backdoor (or targeted poisoning) attacks. These attacks are initiated by malicious clients who seek to compromise the learning process by introducing specific behaviors into the learned model that can be triggered by carefully crafted inputs. Existing FL safeguards have various limitations: They are restricted to specific data distributions or reduce the global model accuracy due to excluding benign models or adding noise, are vulnerable to adaptive defense-aware adversaries, or require the server to access local models, allowing data inference attacks. This paper presents a novel defense mechanism, CrowdGuard, that effectively mitigates backdoor attacks in FL and overcomes the deficiencies of existing techniques. It leverages clients' feedback on individual models, analyzes the behavior of neurons in hidden layers, and eliminates poisoned models through an iterative pruning scheme. CrowdGuard employs a server-located stacked clustering scheme to enhance its resilience to rogue client feedback. The evaluation results demonstrate that CrowdGuard achieves a 100% True-Positive-Rate and True-Negative-Rate across various scenarios, including IID and non-IID data distributions. Additionally, CrowdGuard withstands adaptive adversaries while preserving the original performance of protected models. To ensure confidentiality, CrowdGuard uses a secure and privacy-preserving architecture leveraging Trusted Execution Environments (TEEs) on both client and server sides.

Torsten Krauß, Alexandra Dmitrienko

Federated Learning (FL) facilitates decentralized machine learning model training, preserving data privacy, lowering communication costs, and boosting model performance through diversified data sources. Yet, FL faces vulnerabilities such as poisoning attacks, undermining model integrity with both untargeted performance degradation and targeted backdoor attacks. Preventing backdoors proves especially challenging due to their stealthy nature. Prominent mitigation techniques against poisoning attacks rely on monitoring certain metrics and filtering malicious model updates. While shown effective in evaluations, we argue that previous works didn't consider realistic real-world adversaries and data distributions. We define a new notion of strong adaptive adversaries, capable of adapting to multiple objectives simultaneously. Through extensive empirical tests, we show that existing defense methods can be easily circumvented in this adversary model. We also demonstrate, that existing defenses have limited effectiveness when no assumptions are made about underlying data distributions. We introduce Metric-Cascades (MESAS), a novel defense method for more realistic scenarios and adversary models. MESAS employs multiple detection metrics simultaneously to identify poisoned model updates, creating a complex multi-objective optimization problem for adaptive attackers. In our extensive evaluation featuring nine backdoors and three datasets, MESAS consistently detects even strong adaptive attackers. Furthermore, MESAS outperforms existing defenses in distinguishing backdoors from data distribution-related distortions within and across clients. MESAS is the first defense robust against strong adaptive adversaries, effective in real-world data scenarios, with an average overhead of just 24.37 seconds.

Torsten Krauß, Jasper Stang, Alexandra Dmitrienko

Due to costly efforts during data acquisition and model training, Deep Neural Networks (DNNs) belong to the intellectual property of the model creator. Hence, unauthorized use, theft, or modification may lead to legal repercussions. Existing DNN watermarking methods for ownership proof are often non-intuitive, embed human-invisible marks, require trust in algorithmic assessment that lacks human-understandable attributes, and rely on rigid thresholds, making it susceptible to failure in cases of partial watermark erasure. This paper introduces ClearMark, the first DNN watermarking method designed for intuitive human assessment. ClearMark embeds visible watermarks, enabling human decision-making without rigid value thresholds while allowing technology-assisted evaluations. ClearMark defines a transposed model architecture allowing to use of the model in a backward fashion to interwove the watermark with the main task within all model parameters. Compared to existing watermarking methods, ClearMark produces visual watermarks that are easy for humans to understand without requiring complex verification algorithms or strict thresholds. The watermark is embedded within all model parameters and entangled with the main task, exhibiting superior robustness. It shows an 8,544-bit watermark capacity comparable to the strongest existing work. Crucially, ClearMark's effectiveness is model and dataset-agnostic, and resilient against adversarial model manipulations, as demonstrated in a comprehensive study performed with four datasets and seven architectures.

Torsten Krauß, Hamid Dashtbani, Alexandra Dmitrienko

Machine learning is advancing rapidly, with applications bringing notable benefits, such as improvements in translation and code generation. Models like ChatGPT, powered by Large Language Models (LLMs), are increasingly integrated into daily life. However, alongside these benefits, LLMs also introduce social risks. Malicious users can exploit LLMs by submitting harmful prompts, such as requesting instructions for illegal activities. To mitigate this, models often include a security mechanism that automatically rejects such harmful prompts. However, they can be bypassed through LLM jailbreaks. Current jailbreaks often require significant manual effort, high computational costs, or result in excessive model modifications that may degrade regular utility. We introduce TwinBreak, an innovative safety alignment removal method. Building on the idea that the safety mechanism operates like an embedded backdoor, TwinBreak identifies and prunes parameters responsible for this functionality. By focusing on the most relevant model layers, TwinBreak performs fine-grained analysis of parameters essential to model utility and safety. TwinBreak is the first method to analyze intermediate outputs from prompts with high structural and content similarity to isolate safety parameters. We present the TwinPrompt dataset containing 100 such twin prompts. Experiments confirm TwinBreak's effectiveness, achieving 89% to 98% success rates with minimal computational requirements across 16 LLMs from five vendors.