Research map / Label-efficient and robust learning / Recognition and long-tail

MTL-UE: Learning to Learn Nothing for Multi-Task Learning

ICML 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Group-robust classification · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MTL-UE, a plug-and-play generator-based framework that creates unlearnable examples for multi-task datasets using learnable class-wise embeddings plus intra-/inter-task embedding regularization, effective against MTL and STL models.

Problem

Existing unlearnable example methods target single-task learning; protecting multi-task datasets against unauthorized training of multi-task learning (MTL) models, including dense prediction tasks, remains largely unexplored.

Approach

Replace per-sample perturbation optimization with an encoder-decoder generator conditioned on learnable task-specific class-wise feature embeddings (label priors), regularized by intra-task cosine dispersion and inter-task independence losses, so that it can plug into surrogate-dependent UE methods (EM, TAP, SEP).

Key results

• On CelebA with ResNet-18 MTL model, MTL-UE-SEP reduces average accuracy to 58.73% (vs 84.25% for SEP) and MTL-UE-TAP to 59.51% (vs 85.24% for TAP).
• On NYUv2 MTL segmentation, MTL-UE-EM reduces mIoU to 2.37 (vs 26.45 for EM), and MTL-UE-SEP achieves the best normal estimation error (Mean 41.78).
• MTL-UE-TAP achieves 68.65% average accuracy on STL CelebA models, versus 87.00% for TAP.

Measured effects reported in the paper

Surprising findings

• Surrogate-dependent baselines (EM, TAP, SEP) underperform patch-based surrogate-free AR on MTL models, attributed to uncontrolled intra-class variance.
• Partial data protection is ineffective: mixing even 20% clean data noticeably weakens the protection on NYUv2.

Limitations

• Performance on STL models degrades as the number of tasks grows beyond about 15 tasks on CelebA; authors leave this for future work.
• Partial data protection is ineffective; effectiveness drops quickly when data is not fully unlearnable.
• Results on ChestX-ray14 for TAP and SEP are weak because the clean surrogate model performs poorly.

Experimental setup

Datasets: CelebA ChestX-ray14 UTKFace NYUv2

Benchmarks: CelebA ChestX-ray14 UTKFace NYUv2

Metrics: accuracy AUC-ROC mIoU pixel accuracy absolute error relative error mean absolute error median absolute error

Baselines: EM TAP SEP LSP AR

Builds on: unlearnable examples (error-minimizing noise) targeted adversarial poisoning (TAP) SEP autoregressive perturbations (AR) hard parameter sharing (HPS)

Closest papers by meaning

• Distribution Matching for Multi-Task Learning of Classification Tasks: A Large-Scale Study on Faces & Beyond (AAAI 2024 · similarity 0.87)
• NTKMTL: Mitigating Task Imbalance in Multi-Task Learning from Neural Tangent Kernel Perspective (NeurIPS 2025 · similarity 0.87)
• MTLoRA: Low-Rank Adaptation Approach for Efficient Multi-Task Learning (CVPR 2024 · similarity 0.87)
• Towards Consistent Multi-Task Learning: Unlocking the Potential of Task-Specific Parameters (CVPR 2025 · similarity 0.87)
• Identifying and Mitigating Spurious Correlation in Multi-Task Learning (CVPR 2025 · similarity 0.87)
• AdaMerging: Adaptive Model Merging for Multi-Task Learning (ICLR 2024 · similarity 0.86)