Multimodal Detection Architecture
Our mission is to secure the digital ecosystem against synthetic manipulation. ZeroTrue employs an ensemble of evidence-first models to detect AI-generated content across text, code, voice, music, and video.
Architectures by Modality
Our specialized pipelines are designed to capture the unique artifacts left by generative models in each domain.
2.1 Generated Text
Technical Approach
Hybrid detectors combining likelihood/perplexity tests, supervised classifiers (RoBERTa/DeBERTa fine-tunes), and stylometry features.
Key Techniques
2.2 LLM-Generated Code
Technical Approach
Dual-track analysis using content-agnostic detectors (AST/CFG features) and provenance-aware watermark readers.
Key Techniques
2.3 Voice Cloning / TTS
Technical Approach
Spectrogram-level CNN/Conformer models fused with LFCC/EFCC anti-spoof features and ECAPA-TDNN variants.
Key Techniques
2.4 Generated Music
Technical Approach
Multi-scale spectrogram fingerprints and timbre/harmonic residuals, ensembled with music-theory features.
Key Techniques
2.5 Deepfake Video
Technical Approach
Spatiotemporal detectors (Xception/EfficientNet + TimeSformer) with frequency and physiology auxiliaries.
Key Techniques
Datasets & Evaluation Protocols
| Dataset | Domain | Size / Notes | Metric(s) |
|---|---|---|---|
| HC3 / HC3+ | Text | High-quality ChatGPT vs Human | AUROC, FPR@TPR |
| RAID | Text | Adversarial attacks & domains | Robustness Score |
| AIGCodeSet | Code | Python generation tasks | AUROC |
| ASVspoof 2019/21 | Audio | Logical/Physical Access | EER, min t-DCF |
| ADD 2022 | Audio | Audio Deepfake Detection | EER |
| DFDC | Video | 100k+ clips, Facebook backed | Video-AUC |
| FaceForensics++ | Video | Diverse manipulation methods | Frame-AUC |
| Celeb-DF | Video | High-quality Deepfakes | AUC |
* We also track GenImage for image generation baselines.
Evaluation Metrics
Operational Bands
Recommended action thresholds based on confidence scores.
Provenance & Standards
ZeroTrue complements statistical detection with cryptographic provenance. Where available, we parse and display C2PA Content Credentials alongside our detector verdicts. This provides a dual-layer defense: verifying the "authentic" chain of custody while statistically flagging likely manipulations.
Selected References
Research Roadmap
Multi-model Attribution
Fingerprinting specific generator versions (e.g., Midjourney v6 vs DALL-E 3).
Watermark-Aware Fusion
Integrating hidden watermarks into confidence scoring.
Adversarial Hardening
Post-training robustness against new evasion attacks.