The Replay Tax: auditing the privacy cost of continual-learning strategies in object detection — per-image attack scores, splits, buffers and analysis code
Year: 2026
Type: dataset
Abstract: This archive is the replication package for a membership-inference audit of continual-learning strategies in object detection. Six training arms (naive, no adaptation, joint, replay, elastic weight consolidation, learning without forgetting) are evaluated across two datasets and two detector architectures with three seeds each, under three adversaries of increasing realism: a loss threshold, a difficulty-calibrated attack against a reference model, and a label-free attack that observes only detection confidences. The primary cell additionally carries a likelihood-ratio attack with 32 shadow models. The archive contains the per-image attack scores for every arm, seed and cell; the member, non-member, control and validation-fitting split listings; the exact buffer contents for every buffer size and seed; per-arm utility measurements; the likelihood-ratio scores and shadow-membership tables; and all analysis code, including the scripts that generate every table and figure in the paper. The images themselves are not redistributed. Both corpora are third-party public research datasets (a weapon-detection corpus on Roboflow Universe and the Teledyne FLIR Free ADAS Thermal Dataset), and the archive supplies the file lists needed to reconstruct our exact splits from them. The replay buffers are additionally reproducible from their seed alone. A verification script, independent of the analysis code, recomputes every headline number in the paper from the raw per-image scores and reports 37 of 37 claims reproduced. It runs on CPU in about a minute and needs no images and no GPU. The package distinguishes throughout between two buffer constructions, because they give different answers: an audit buffer drawn from the audited member set, and a deployment buffer drawn from the whole first task with no reference to the audit. Arms of the second kind are tagged replayU.
Institution: Sivas Cumhuriyet Üniversitesi
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204906463