ML Security PapersPromptScreen: Efficient Jailbreak Mitigation Using Semantic Linear Classification in a Multi-Staged Pipeline
Akshaj Prashanth Rao 1, Advait Singh 1, Saumya Kumaar Saksena 1, Dhruv Kumar 1,2
Published on arXiv
2512.19011
Prompt Injection
OWASP LLM Top 10 — LLM01
Key Finding
SVM-based configuration achieves 93.4% accuracy and reduces average time-to-completion from ~450s to 47s, yielding over 10x lower latency than ShieldGemma while improving accuracy from 35.1% to 93.4%.
PromptScreen
Novel technique introduced
Prompt injection and jailbreaking attacks pose persistent security challenges to large language model (LLM)-based systems. We present PromptScreen, an efficient and systematically evaluated defense architecture that mitigates these threats through a lightweight, multi-stage pipeline. Its core component is a semantic filter based on text normalization, TF-IDF representations, and a Linear SVM classifier. Despite its simplicity, this module achieves 93.4% accuracy and 96.5% specificity on held-out data, substantially reducing attack throughput while incurring negligible computational overhead. Building on this efficient foundation, the full pipeline integrates complementary detection and mitigation mechanisms that operate at successive stages, providing strong robustness with minimal latency. In comparative experiments, our SVM-based configuration improves overall accuracy from 35.1% to 93.4% while reducing average time-to-completion from approximately 450 s to 47 s, yielding over 10 times lower latency than ShieldGemma. These results demonstrate that the proposed design simultaneously advances defensive precision and efficiency, addressing a core limitation of current model-based moderators. Evaluation across a curated corpus of over 30,000 labeled prompts, including benign, jailbreak, and application-layer injections, confirms that staged, resource-efficient defenses can robustly secure modern LLM-driven applications.
Key Contributions
- Lightweight semantic filter using TF-IDF representations and a Linear SVM classifier that achieves 93.4% accuracy and 96.5% specificity with negligible computational overhead
- Multi-stage defense pipeline integrating heuristic filters, semantic classification, cluster-based similarity detection, vectorized attack memory retrieval, and model-based sequence assessment
- Curated evaluation corpus of 30,000+ labeled prompts covering benign, jailbreak, and application-layer injection samples with reproducible evaluation framework