Calibrated membership-inference attack against pretraining data. Uses per-token log-likelihood vs. neighborhood entropy to achieve state-of-the-art AUROC across LLaMA, Pythia, GPT-NeoX families. Decouples token informativeness from membership signal, producing a clean p-value under permutation null.

Detection of copyrighted content via paraphrase-distinguishing probes. Forces the model to choose between original passages and high-quality paraphrases; preference for the original is statistically attributable to memorization.

Non-verbatim recall measurement. Replaces exact-match with semantic-equivalence scoring under a calibrated neighborhood, recovering memorization signal that survives stylistic rewriting and translation.

Probabilistic discoverable extraction with confidence bounds. Estimates the probability that a target string is recoverable from a model under a budgeted prompt distribution, with formal lower-bound guarantees.

Training-data attribution at scale via random-projection of gradients. Returns a per-training-example influence score on a target query.

Scalable TRAK variant for billion-parameter diffusion models. Uses distilled gradient surrogates and locality-sensitive hashing for tractable attribution across catalog-scale training indices.

Low-rank adapter probes that fingerprint memorization-prone parameter subspaces. Surface circuits where training-data residue concentrates.

Speech and music MIA combining acoustic loss-curvature, mel-spectrogram divergence, and waveform-level NN-distance.

Contrastive style descriptors capturing artist-specific visual fingerprints invariant to subject matter. Validated for Getty / artist-style cases.

Melodic similarity across pitch-class profiles and rhythmic envelopes for music memorization claims.

Spatiotemporal action-level latent localization for short-form video memorization detection.

Forensic detection of model merges and lineage via centered kernel alignment between candidate parent models and a target.

Fisher / Stouffer / Simes / harmonic-mean / Cauchy combiners over independent MIA detectors. Patentable ensemble — top novelty rating.

Collective dataset inference at the concept level for image distributions.

Two-stage detector that disentangles RAG-retrieval recall from parametric memorization.

Two-sample tests (MMD, Kolmogorov–Smirnov, energy distance) for concept-level recall in generative models.

Diffusion-model attribution via inverted embedding probes. Recovers a pseudo-token that best activates target style; statistical significance of activation vs. distractors yields attribution score. Patentable invention, top novelty rating.

Self-supervised copy detection embeddings combined with DINOv2 features for image-level provenance and near-duplicate detection.

Six-axis verifier for unlearning and deletion claims: verbatim, knowledge, privacy, bias, utility, and scalability.

Tight empirical lower bound on the differential-privacy epsilon of a deployed model via membership-inference adversary.