Embedded Files
The Genμ Challenge @ U&Me 2025 Workshop
Challenge Description: As part of the workshop, we are organizing the Generative Machine Unlearning (Genμ) Challenge to benchmark and advance research in generative machine unlearning and model editing by providing a structured framework to evaluate forgetting effectiveness, retention accuracy, and robustness against adversarial manipulations. The Kaggle link to the challenge is NOW OPEN
Task Description: The goal is to remove the target concept from the vision foundation models (text-to-image generative models) while preserving the model’s overall functionality and coherence. The resulting “unlearned” model should also demonstrate robustness against both engineered and adversarial prompts intended to reintroduce or exploit the erased concept.
Dataset: The competition will introduce a benchmark dataset specifically designed for generative unlearning. It will consist of three key components:
Forget Set: includes concepts designated for removal.
Locality Set: contains unrelated concepts that must remain unaffected
Adjacency Set: assesses unintended alterations in closely related concepts.
Evaluation Methodology: Submissions will be evaluated based on their effectiveness in erasing target concepts while preserving both general and related knowledge with minimal effect on the model weights. The Forgetting-Retention Score (FRS) will serve as the primary evaluation metric, balancing forgetting accuracy---the effectiveness of concept removal---with retention accuracy, which measures how well related and unrelated concepts are maintained. Additionally, robustness against engineered and adversarial prompts will be tested to ensure erased concepts do not reappear. Perceptual and distributional stability will also be analyzed to verify consistent model behavior post-unlearning. To facilitate comparison, baseline methods will be provided.
Challenge Deadline: August 1, 2025. We plan to publish a full paper on the challenge and its results. To encourage participation, we will invite the three highest-scoring teams to join the writing effort and be listed as co-authors, provided they outperform the baseline.
Check out more information at: https://unlearning.iab-rubric.org
Kaggle Challenge link: https://www.kaggle.com/competitions/gen-challenge-u-me-workshop-iccv-25/overview
Page updated
Google Sites
Report abuse