Research Citations
Every model, method, and dataset underpinning Aiscern is grounded in peer-reviewed research. We cite all work we build on — transparency is a core part of how we operate.
Ensemble Accuracy Benchmarks
Measured against held-out evaluation sets per modality. Full methodology on the methodology page.
AI Text Verification
Hans et al. · ICML 2024
Cross-perplexity scoring between two observer LLMs — the core of our text verification signal.
Liu et al. · arXiv 2019
Backbone for the openai-detector fine-tune used in our text ensemble.
Koike et al. · ACL 2024
Evaluation of perplexity-based zero-shot detectors — informed our threshold calibration.
Wang et al. · EACL 2024
Benchmark dataset and methodology for cross-generator verification — used in our text benchmarks.
Dugan et al. · ACL 2024
Adversarial robustness evaluation framework — used to stress-test our text models.
Deepfake Image Verification
Dosovitskiy et al. · ICLR 2021
Foundation architecture for our image verification classifier.
Radford et al. (OpenAI) · ICML 2021
Embedding model for our image RAG pipeline and similarity-based retrieval.
Ojha et al. · CVPR 2023
Zero-shot generalization approach for verifying unseen generators — informs our ensemble strategy.
Corvi et al. · ICASSP 2023
Frequency-domain forensics for diffusion model outputs — basis for our spectral signal layer.
AI Audio & Voice Clone Verification
Baevski et al. · NeurIPS 2020
Foundation model for our audio deepfake verification fine-tune.
Yamagishi et al. · ASVspoof 2021
Primary evaluation dataset and challenge protocol for our audio models.
Yi et al. · ICASSP 2023
Extended benchmark including emotional and singing deepfakes.
Chen et al. · Interspeech 2022
Spectrogram fusion approach informing our multi-feature audio pipeline.
Video / Deepfake Verification
Rössler et al. · ICCV 2019
Primary training and evaluation dataset for face manipulation verification.
Zheng et al. · arXiv 2021
ViT-based frame-level deepfake detector used in our video ensemble.
Agarwal et al. · IEEE Workshops 2020
Temporal behavioral signals for video deepfake verification.
RAG & Ensemble Methods
Lewis et al. · NeurIPS 2020
Foundational RAG paper — architecture behind our vector-retrieval confidence augmentation.
Mitchell et al. · FAccT 2019
Guidelines we follow for documenting bias, limitations, and intended use of all models.
pgvector contributors · GitHub
Vector database extension used for pattern storage and cosine similarity retrieval.
Missing a citation?
If you believe we are using your work without attribution, please contact research@aiscern.com and we will add it promptly.
Also see: Benchmarks · Methodology