ML Security PapersNo Trust Issues Here: A Technical Report on the Winning Solutions for the Rayan AI Contest
Ali Nafisi 1, Sina Asghari 2, Mohammad Saeed Arvenaghi 2, Hossein Shakibania 3
Published on arXiv
2512.01498
Model Poisoning
OWASP ML Top 10 — ML10
Key Finding
Backdoor detection method achieves 78% accuracy on hidden trigger identification, contributing to an overall 1st-place finish in the Rayan AI Trustworthy AI Contest.
This report presents solutions to three machine learning challenges developed as part of the Rayan AI Contest: compositional image retrieval, zero-shot anomaly detection, and backdoored model detection. In compositional image retrieval, we developed a system that processes visual and textual inputs to retrieve relevant images, achieving 95.38% accuracy and ranking first with a clear margin over the second team. For zero-shot anomaly detection, we designed a model that identifies and localizes anomalies in images without prior exposure to abnormal examples, securing second place with a 73.14% score. In the backdoored model detection task, we proposed a method to detect hidden backdoor triggers in neural networks, reaching an accuracy of 78%, which placed our approach in second place. These results demonstrate the effectiveness of our methods in addressing key challenges related to retrieval, anomaly detection, and model security, with implications for real-world applications in industries such as healthcare, manufacturing, and cybersecurity. Code for all solutions is available online (https://github.com/safinal/rayan-ai-contest-solutions).
Key Contributions
- Backdoor detection method achieving 78% accuracy at identifying hidden triggers in neural networks
- Compositional image retrieval pipeline (Token Classification + Embedding Arithmetic) achieving 95.38% accuracy
- Zero-shot anomaly detection system achieving 73.14% score, securing 2nd place
🛡️ Threat Analysis
The backdoored model detection track proposes a method to identify hidden backdoor triggers in neural networks — a direct defense against model poisoning/trojan attacks.