SeedPrints: Tracking the Training Seed of Large Language Models via Fingerprints

SeedPrints is a groundbreaking model provenance technology that generates unique "fingerprints" by analyzing model outputs, enabling accurate identification of the random seeds used in training large language models, and providing new means for AI security and model auditing. This article will introduce the technology's background, principles, experimental validation, application value, and future directions.

Security Challenges in Model Provenance

With the widespread application of large language models in various fields, model security and auditability have become focal points. Traditional model identification methods focus on architecture and parameters, but the critical information of training random seeds is often regarded as an untraceable black box.

Research Background and Core Findings

SeedPrints is a research work accepted by ICLR 2026. The core hypothesis is: The random seeds used in training large language models leave unique "fingerprints" in their behavioral patterns, which can be extracted and identified through carefully designed detection methods. Even if models have the same architecture, dataset, and hyperparameters, models with different seeds will exhibit distinguishable features under specific test conditions.

SeedPrints' technical implementation includes three key steps:

1. Fingerprint Extraction: Design a series of probe tasks, observe model response patterns to capture behavioral features, and maximize the difference signals between models with different seeds.
2. Fingerprint Encoding: Convert the extracted features into high-dimensional vector representations to ensure the stability and distinguishability of the fingerprints.
3. Seed Identification: Compare the fingerprint of the model to be detected with the fingerprint library of models with known seeds using machine learning classifiers or similarity matching algorithms to infer the training seed.

Experimental results show that SeedPrints can identify training seeds with high accuracy across multiple mainstream large language model architectures. Even after models are fine-tuned or quantized, the fingerprints remain robust.

The research team discussed factors affecting accuracy: Fingerprints of larger models are more stable, and carefully designed combinations of probe tasks can significantly improve recognition performance, providing guidance for parameter selection in practical applications.

Positive Implications

• Model Auditing and Provenance: Helps identify unauthorized model copying or theft.
• Open Source Ecosystem Validation: Ensures that distributed models come from the claimed training process.

Security Concerns

Attackers may use fingerprints to infer training details, providing new avenues for model theft or adversarial attacks. The community needs to explore how to balance legitimate auditing and preventing abuse.

The SeedPrints open-source code repository provides a complete implementation, including modules for data preparation, probe task execution, feature extraction, fingerprint generation, and seed identification.

Users need to prepare the model to be detected and reference models with known seeds, and complete fingerprint extraction and comparison through API interfaces. The project supports custom probe tasks and encoding schemes to adapt to different scenarios, and provides detailed documentation and example scripts.

Limitations

• The accuracy of fingerprint identification is affected by multiple factors, which may lead to misjudgments.
• Specific training techniques or post-processing methods may reduce the effectiveness of fingerprints.

Future Directions

• Develop more robust fingerprint extraction methods.
• Explore scalability for larger-scale models.
• Study the relationship between fingerprints and model capabilities/biases.
• Establish a complete standard framework for model provenance.