Inspiration
What if the very photograph taken to protect an animal could become the information that puts its life at risk? Wildlife imagery is increasingly important for research and conservation, but a single image can unintentionally expose sensitive locations through metadata, visible text, captions, or other hidden information. For endangered species, such exposure could potentially contribute to poaching, illegal hunting, or targeted exploitation. WildGuard was created to protect both the evidence and the wildlife behind it.
What it does
WildGuard accepts a wildlife image and first evaluates it through a 47-feature XGBoost risk pipeline. It then performs deeper analysis using OCR and deterministic security checks to identify sensitive information and explain exactly why an image has been flagged. Based on the resulting risk, WildGuard can automatically sanitize the image by removing sensitive metadata and location information and redacting exposed visual content.
The protected image is then given a cryptographic provenance record using SHA-256 and embedded directly into the artifact. Through the Version interface, that provenance can later be verified. If the artifact remains unchanged, WildGuard reports MATCHED. If it has been modified or tampered with, the system detects the discrepancy and flags the variation.
How we built it
WildGuard was built as a modular local application combining XGBoost, Tesseract OCR, image and metadata processing, lightweight steganographic analysis, a policy engine, automated sanitization, SHA-256 hashing, and byte-exact provenance embedding.
The analysis pipeline combines five categories of signals into the 47-feature vector used consistently for training and inference. The system deliberately uses two detection layers: machine learning identifies statistical patterns, while deterministic checks provide direct evidence such as detected GPS or camera-trap information.
This ensures that a critical security signal is not overlooked simply because the model predicts a low risk.
Challenges we ran into
The biggest challenge was making the system robust without relying on a single detection mechanism. OCR can miss difficult text, while machine learning can identify patterns without always providing direct evidence. WildGuard therefore combines both approaches while keeping their roles distinct.
Provenance presented another challenge. Re-encoding an image can change its bytes, making ordinary hashing unreliable. WildGuard solves this through byte-exact insertion and extraction, allowing the pre-provenance artifact to be recovered and verified without relying on a lossy re-encode.
Accomplishments that we're proud of
WildGuard became a complete end-to-end system rather than just a classifier. It can analyse an image, explain its risks, sanitize it, establish provenance, and subsequently verify its integrity.
The documented evaluation achieved 90.5% test accuracy, 95.2% ROC-AUC, 91.1% precision, and 86.0% recall, while the mixed stress test achieved 100% across its reported metrics.
The implementation also includes 39 automated tests covering metadata, OCR, filenames, steganography, sanitization, hashing, and provenance.
What we learned
We learned that protecting sensitive environmental data requires more than a good model. Machine learning can identify patterns, but security decisions also need concrete evidence, transparent explanations, reliable remediation, and verifiable integrity.
We also learned the importance of designing for reproducibility and data-leakage resistance rather than optimizing only for a single impressive metric.
What's next for WildGuard
The next step is validation on the full real-world dataset rather than the synthetic development replica used for the documented evaluation. The project can then be strengthened through broader real-world testing, improved OCR robustness, expanded species-sensitivity knowledge, and further refinement of the detection and provenance systems.
WildGuard has already established the complete workflow from detection and analysis through sanitization, provenance creation, and verification. The next stage is taking that foundation into real-world conservation environments.
Log in or sign up for Devpost to join the conversation.