How DeepGuard
Actually Works
Our detection system combines six independent analysis layers, trained on over 2 million labeled samples across images, video, audio, and text. Here's the full technical picture — no marketing fluff.
Six Detection Layers
Each layer catches a different class of manipulation. Results are cross-validated — a single layer can't produce a false verdict.
Pixel-Level Forensics
Analyzes individual pixel neighborhoods for statistical anomalies introduced by diffusion models and GANs. Detects upsampling artifacts, frequency domain inconsistencies, and unnatural noise patterns.
Multi-Model Ensemble
Runs 4 independently trained detection models in parallel and cross-validates results. No single model failure can produce a false verdict — all models must agree for high-confidence outputs.
Semantic Consistency Check
Detects logical inconsistencies that AI models commonly produce: impossible lighting, anatomical errors, background incoherence, and text rendering failures.
Metadata & Provenance
Inspects EXIF data, file structure, and modification history. AI-generated files typically lack camera metadata, have inconsistent timestamps, or show signs of post-processing software.
Temporal Analysis (Video)
For video content, analyzes inter-frame consistency, temporal coherence of facial features, and blending boundary artifacts that appear at face-swap edges across frames.
Prosody & Voice Forensics
Audio analysis detects TTS synthesis artifacts: unnatural prosody patterns, spectral inconsistencies, missing breath sounds, and the characteristic "flatness" of AI-generated speech.
Accuracy by AI Model
Tested on held-out validation sets not seen during training. Numbers reflect detection accuracy on real-world content, not curated lab samples.
Benchmarks conducted Q4 2024. Results may vary on highly compressed or low-resolution content. Full methodology available on request.
Training Datasets
Models are trained on publicly available academic datasets plus proprietary collections. We do not train on user-submitted content.
Privacy & Data Handling
Zero Data Retention
Uploaded files are processed in isolated containers and deleted immediately after analysis. We never store your content.
Encrypted Transit
All file transfers use TLS 1.3. Files are processed in memory where possible and never written to persistent disk.
No Training on User Data
We explicitly do not use submitted content to improve or retrain our models. Your data is yours.
Audit Logs
Enterprise customers receive full audit logs of all API calls, analysis results, and access events for compliance.