Addressing the Devastating Effects of Single-Task Data Poisoning in Exemplar-free Continual Learning

Stanisław Pawlak, Bartłomiej Twardowski, Tomasz Trzcinski, Joost van de Weijer
Proceedings of The 4th Conference on Lifelong Learning Agents, PMLR 330:162-181, 2026.

Abstract

Our research addresses the overlooked security concerns related to data poisoning in continual learning (CL). Data poisoning – the intentional manipulation of training data to affect the predictions of machine learning models – was recently shown to be a threat to CL training stability. While existing literature predominantly addresses scenario-dependent attacks, we propose to focus on a more simple and realistic single-task poison (STP) threats. In contrast to previously proposed poisoning settings, in STP adversaries lack knowledge and access to the model, as well as to both previous and future tasks. During an attack, they only have access to the current task within the data stream. Our study demonstrates that even within these stringent conditions, adversaries can compromise model performance using standard image corruptions. We show that STP attacks are able to strongly disrupt the whole continual training process: decreasing both the stability (its performance on past tasks) and plasticity (capacity to adapt to new tasks) of the algorithm. Finally, we propose a high-level defense framework for CL along with a poison task detection method based on task vectors.

Cite this Paper

BibTeX
@InProceedings{pmlr-v330-pawlak26a, title = {Addressing the Devastating Effects of Single-Task Data Poisoning in Exemplar-free Continual Learning}, author = {Pawlak, Stanis{\l}aw and Twardowski, Bart{\l}omiej and Trzcinski, Tomasz and Weijer, Joost van de}, booktitle = {Proceedings of The 4th Conference on Lifelong Learning Agents}, pages = {162--181}, year = {2026}, editor = {Chandar, Sarath and Pascanu, Razvan and Eaton, Eric and Liu, Bing and Mahmood, Rupam and Rannen-Triki, Amal}, volume = {330}, series = {Proceedings of Machine Learning Research}, month = {11--14 Aug}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v330/main/assets/pawlak26a/pawlak26a.pdf}, url = {https://proceedings.mlr.press/v330/pawlak26a.html}, abstract = {Our research addresses the overlooked security concerns related to data poisoning in continual learning (CL). Data poisoning – the intentional manipulation of training data to affect the predictions of machine learning models – was recently shown to be a threat to CL training stability. While existing literature predominantly addresses scenario-dependent attacks, we propose to focus on a more simple and realistic single-task poison (STP) threats. In contrast to previously proposed poisoning settings, in STP adversaries lack knowledge and access to the model, as well as to both previous and future tasks. During an attack, they only have access to the current task within the data stream. Our study demonstrates that even within these stringent conditions, adversaries can compromise model performance using standard image corruptions. We show that STP attacks are able to strongly disrupt the whole continual training process: decreasing both the stability (its performance on past tasks) and plasticity (capacity to adapt to new tasks) of the algorithm. Finally, we propose a high-level defense framework for CL along with a poison task detection method based on task vectors.} }
Endnote
%0 Conference Paper %T Addressing the Devastating Effects of Single-Task Data Poisoning in Exemplar-free Continual Learning %A Stanisław Pawlak %A Bartłomiej Twardowski %A Tomasz Trzcinski %A Joost van de Weijer %B Proceedings of The 4th Conference on Lifelong Learning Agents %C Proceedings of Machine Learning Research %D 2026 %E Sarath Chandar %E Razvan Pascanu %E Eric Eaton %E Bing Liu %E Rupam Mahmood %E Amal Rannen-Triki %F pmlr-v330-pawlak26a %I PMLR %P 162--181 %U https://proceedings.mlr.press/v330/pawlak26a.html %V 330 %X Our research addresses the overlooked security concerns related to data poisoning in continual learning (CL). Data poisoning – the intentional manipulation of training data to affect the predictions of machine learning models – was recently shown to be a threat to CL training stability. While existing literature predominantly addresses scenario-dependent attacks, we propose to focus on a more simple and realistic single-task poison (STP) threats. In contrast to previously proposed poisoning settings, in STP adversaries lack knowledge and access to the model, as well as to both previous and future tasks. During an attack, they only have access to the current task within the data stream. Our study demonstrates that even within these stringent conditions, adversaries can compromise model performance using standard image corruptions. We show that STP attacks are able to strongly disrupt the whole continual training process: decreasing both the stability (its performance on past tasks) and plasticity (capacity to adapt to new tasks) of the algorithm. Finally, we propose a high-level defense framework for CL along with a poison task detection method based on task vectors.
APA
Pawlak, S., Twardowski, B., Trzcinski, T. & Weijer, J.v.d.. (2026). Addressing the Devastating Effects of Single-Task Data Poisoning in Exemplar-free Continual Learning. Proceedings of The 4th Conference on Lifelong Learning Agents, in Proceedings of Machine Learning Research 330:162-181 Available from https://proceedings.mlr.press/v330/pawlak26a.html.

Related Material