analiza aiqbench
This paper is squarely about AI: it studies memorization in diffusion models, a core machine learning topic, and proposes statistical monitoring methods to detect when a generative model reproduces training data. The contribution is methodological and modest in scope, evaluated on a single unconditional CIFAR-10 setup, so its results should be read as a proof of concept rather than a general solution. For readers, the practical takeaway is that adaptive, trajectory-level detection can catch memorization earlier and with less conservative false-alarm control than fixed-timestep checks, which matters for auditing generative models for privacy and copyright risk.
Valid Early Detection of Memorization in Diffusion Models
arXiv:2610.34617v1 Announce Type: new Abstract: Diffusion models can sometimes reproduce examples from their training data, a phenomenon known as memorization. Existing memorization detection methods for unconditional diffusion models typically operate at a pre-selected timestep and therefore cannot adapt the detection time to the evolving denoising trajectory. Sequential monitoring enables adaptive detection but repeated testing across timesteps complicates trajectory-level false-alarm control. We study two ways of combining split conformal calibration with sequential monitoring for prompt-free memorization detection: per-step Bonferroni calibration and direct trajectory-level calibration of the maximum memorization score. The former controls repeated testing by splitting the false-alarm budget across timesteps, whereas the latter calibrates the event of ever crossing a threshold directly through the trajectory-level maximum. We evaluate this formulation on an unconditional CIFAR-10 EDM using two complementary per-step statistics based on geometry and representation spikiness. Experiments show that sequential monitoring improves the power-timing tradeoff over fixed-time detection. Compared with Bonferroni calibration, trajectory-level calibration achieves comparable detection power while using the false-alarm budget less conservatively and enabling earlier detection at higher noise levels.
Abstract:Diffusion models can sometimes reproduce examples from their training data, a phenomenon known as memorization. Existing memorization detection methods for unconditional diffusion models typically operate at a pre-selected timestep and therefore cannot adapt the detection time to the evolving denoising trajectory. Sequential monitoring enables adaptive detection but repeated testing across timesteps complicates trajectory-level false-alarm control. We study two ways of combining split conformal calibration with sequential monitoring for prompt-free memorization detection: per-step Bonferroni calibration and direct trajectory-level calibration of the maximum memorization score. The former controls repeated testing by splitting the false-alarm budget across timesteps, whereas the latter calibrates the event of ever crossing a threshold directly through the trajectory-level maximum. We evaluate this formulation on an unconditional CIFAR-10 EDM using two complementary per-step statistics based on geometry and representation spikiness. Experiments show that sequential monitoring improves the power-timing tradeoff over fixed-time detection. Compared with Bonferroni calibration, trajectory-level calibration achieves comparable detection power while using the false-alarm budget less conservatively and enabling earlier detection at higher noise levels.
Extended version of a 4-page paper accepted as a poster at the NeurIPS 2026 Workshop on E-Values: From Statistics to ML. 13 pages, 3 figures, 2 tables
Machine Learning (cs.LG); Machine Learning (stat.ML)
arXiv:2610.34617 [cs.LG]
(or arXiv:2610.34617v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2610.34617 arXiv-issued DOI via DataCite (pending registration)
Text integral sub licența deschisă a publicației · 216 cuvinte
Subiecte conexe
The Mirrored Influence Hypothesis: Efficient Data Influence Estimation by Harnessing Forward Passes
arXiv · 9 iun. 2026
Detecting Functional Memorization in Code Language Models
arXiv · 12 iun. 2026
Detection of Memorization in Diffusion Models Through Thermodynamic Singularity
arXiv · 21 oct. 2026
Mitigating Memorization Where It Happens
arXiv · 16 oct. 2026