Adaptive Attacks Expose SLM Vulnerabilities and Qualitative Insights
6 Feb 2025
Adaptive attacks require larger perturbations to overcome TDNF defenses in SLMs, reducing jailbreak success; qualitative examples highlight strengths and limita
Transfer Attacks Reveal SLM Vulnerabilities and Effective Noise Defenses
6 Feb 2025
Cross-model attacks expose SLM weaknesses, while noise-based defenses substantially reduce jailbreak risks with minimal impact on performance.
Cross-Prompt Attacks and Data Ablations Impact SLM Robustness
6 Feb 2025
Examine how cross-prompt attacks, training data ablations, and random noise influence the robustness, helpfulness, and safety of speech language models.
Safety Alignment and Jailbreak Attacks Challenge Modern LLMs
6 Feb 2025
Explore how safety alignment and adversarial jailbreak attacks expose vulnerabilities in multimodal LLMs and speech language models.
Audio Encoder Pre-training and Evaluation Enhance SLM Safety
6 Feb 2025
Discover our 24-layer Conformer pre-training details and evaluation methods using Claude 2.1 to ensure safety, relevance, and helpfulness in SLMs.
Integrated Speech Language Models Face Critical Safety Vulnerabilities
6 Feb 2025
Adversarial attacks easily bypass safety in SLMs, urging robust defenses and further research to secure multimodal speech-language systems.
SpeechVerse Unites Audio Encoder and LLM for Superior Spoken QA
6 Feb 2025
Discover how SpeechVerse uses a 24-layer Conformer and LLMs like Flan-T5 and Mistral to boost spoken QA performance.
Unified Speech and Language Models Can Be Vulnerable to Adversarial Attacks
6 Feb 2025
Discover how adversarial attacks expose safety gaps in speech language models and how countermeasures can curb jailbreaking risks.
SLMs Outperform Competitors Yet Suffer Rapid Adversarial Jailbreaks
6 Feb 2025
Results show our SLMs outperform public models in safety and relevance but remain highly vulnerable to fast adversarial attacks.