AI Artifact Fingerprinting

Fingerprint every AI artifact
before it leaves your perimeter.

PreserveIP's AI Artifact Fingerprinting engine creates cryptographic and semantic signatures for model weights, training datasets, and prompt logs — then continuously monitors every developer channel to detect unauthorized copies, fine-tunes, or leakage into public repositories.

Request Access → See What We Track
What We Fingerprint

Every AI artifact that can be stolen, copied, or leaked

Traditional DLP tools were built for documents. PreserveIP understands AI artifacts — their structure, naming patterns, size signatures, and semantic content.

.pt / .pth
PyTorch Model Weights
Full and partial checkpoint files. Detects size-matching, hash collisions, and partial weight extraction.
.safetensors
Safetensors Format
HuggingFace-native model format. Fingerprinted at tensor-block level for partial leak detection.
.pkl / .h5
Sklearn / Keras Models
Serialized scikit-learn pipelines and Keras HDF5 files monitored for unauthorized export.
.jsonl / .parquet
Training Datasets
Dataset schema hashing and row-sample fingerprinting detects leakage even of partial dataset exports.
.yaml / config.json
Model Configs
Architecture definitions, hyperparameter files, and tokenizer configs flagged as high-value IP.
prompt logs
Prompt Engineering
System prompts, fine-tuning datasets, and RLHF feedback logs treated as proprietary trade secrets.
Detection Method

Three-layer fingerprinting — impossible to evade

1

Cryptographic Hashing

SHA-256 hash of every registered model file and dataset. Any byte-for-byte copy is caught instantly, regardless of filename or location.

2

Structural Fingerprinting

Layer counts, parameter counts, tensor shapes, and file sizes are profiled. Partial extractions and architecture clones trigger alerts even without hash matches.

3

Semantic Similarity

Claude AI reads commit messages, PR descriptions, and Slack messages for references to internal model names, dataset codenames, and architecture terms — catching disclosure before any file moves.

4

Cross-Platform Correlation

Signals across GitHub, Slack, and Drive are correlated by actor and timeframe — surfacing coordinated exfiltration attempts that no single-channel tool would catch.

Intellectual Property

Patent-pending detection methodology

The AI Artifact Fingerprinting methodology is the subject of a pending patent disclosure co-invented by PreserveIP's founding team at Texas A&M University, covering both the detection architecture and the legal evidence pipeline.

⚙️

2 Pending Patents — AI artifact fingerprinting methodology and model/dataset fingerprinting method, covering cryptographic hashing, structural profiling, and semantic classification of AI IP artifacts. Named inventors: Shreyas Kumar & Srividhya Ragavan, Texas A&M University.

Platform Capabilities

Built for AI security teams, not generic DLP

⚡

30-Second Detection

From the moment a model file appears in a commit or is shared in Slack, classification and alerting complete in under 30 seconds.

🎯

Zero-Copy Architecture

Fingerprints are derived from metadata and sampling — your model files never leave your environment. Privacy-first by design.

📊

Confidence Scoring

Every detection comes with a 0–100% confidence score from Claude AI, reducing false-positive fatigue for your security team.

🔗

Public Repo Monitoring

PreserveIP scans public GitHub repositories for unauthorized fine-tunes or re-uploads of your registered model artifacts.

Get Started

Register your models. We'll watch the rest.

Join design partners getting early access to AI Artifact Fingerprinting.

Request Access → See Model Weight Detection →