Research map / Continual learning / Class-incremental
AAKR: Adversarial Attack-based Knowledge Retention for Continual Semantic Segmentation
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 AAKR, a continual semantic segmentation method that uses targeted adversarial attacks on incoming images, guided by an attack probability module, to make the old model expose forgettable old classes and thereby retain old knowledge without replay memory.
Problem
Catastrophic forgetting in continual semantic segmentation: distillation-based methods underperform replay-based ones, but replay needs extra storage for generative models and extra training time.
Approach
Attack the added images with multi-step targeted adversarial perturbations whose goal labels are allocated per class according to each class's measured forgetfulness (via an attack probability module), so the old model produces richer predictions over old classes that are transferred to the new model through distillation and cross-entropy on a mixed clean-plus-adversarial batch.
Key results
- On Pascal VOC2012 15-1s (6 tasks), AAKR reaches 66.37% mIoU on all classes versus 54.68% for PLOP and 62.70% for RECALL (replay-based).
- On ADE20k 100-10s, AAKR achieves 33.96% all-class mIoU, the best among compared methods.
- Ablation on VOC 15-1s: adversarial attack alone adds +8.29 all-class mIoU over PLOP, and the attack probability module adds a further +3.40.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AAKR instead of RCIL | Continual segmentation model · Pascal VOC2012 19-1 (2 tasks) | mIoU (all) | Helped |
| Use AAKR instead of RCIL | Continual segmentation model · Pascal VOC2012 15-5 (2 tasks) | mIoU (all): +3.11% | Helped |
| Use AAKR instead of RECALL | Continual segmentation model · Pascal VOC2012 15-1s (6 tasks) | mIoU (all) | Helped |
| Use AAKR instead of RCIL | Continual segmentation model · Pascal VOC2012 10-10 (2 tasks) | mIoU (all): +1.36% | Helped |
| Use AAKR instead of RECALL | Continual segmentation model · Pascal VOC2012 10-5s (3 tasks) | mIoU (all): +3.35% | Helped |
| Use AAKR instead of LAG | Continual segmentation model · Pascal VOC2012 10-1s (11 tasks) | mIoU (all) | Hurt |
| Use AAKR instead of RCIL | Continual segmentation model · ADE20k 100-50 (2 tasks) | mIoU (all): +0.66% | Helped |
| Use AAKR instead of LAG | Continual segmentation model · ADE20k 50-50s (3 tasks) | mIoU (all) | Helped |
| Use AAKR instead of LAG | Continual segmentation model · ADE20k 100-10s (6 tasks) | mIoU (all) | Helped |
| Use AAKR instead of LAG | Continual segmentation model · ADE20k 100-5s (11 tasks) | mIoU (all) | Hurt |
| Add adversarial samples using AAKR attack to the PLOP baseline (ablation) | PLOP on Pascal VOC2012 · 15-1s (6 tasks) | mIoU (all): +8.29% | Helped |
| Add APM to construct the attack goal instead of selecting it randomly (ablation) | PLOP with adversarial samples on Pascal VOC2012 · 15-1s (6 tasks) | mIoU (all): +3.40% | Helped |
Surprising findings
- Random-target adversarial samples and generic PGD/SegPGD attacks yield smaller gains or even lower new-class mIoU than the proposed targeted approach (PGD all-class 60.57 vs AAKR 66.37).
Limitations
- Gains on some settings are modest (e.g., 100-50 ADE20k all-class vs RCIL +0.66) and not uniform across settings.
- Requires extra attack iterations (k=3) on added images, adding some training overhead over pure distillation.
Experimental setup
Datasets: Pascal VOC2012 ADE20k
Benchmarks: Pascal VOC2012 ADE20k
Metrics: mIoU (old) mIoU (new) mIoU (all) mIoU (avg)
Baselines: ILT MiB PLOP REMINDER RCIL LAG RECALL PGD SegPGD
Builds on: PLOP PGD SegPGD MiB
Closest papers by meaning
- Attacks on Continual Semantic Segmentation by Perturbing Incremental Samples (AAAI 2024 · similarity 0.88)
- Vanish into Thin Air: Cross-prompt Universal Adversarial Attacks for SAM2 (NeurIPS 2025 · similarity 0.86)
- TIKP: Text-to-Image Knowledge Preservation for Continual Semantic Segmentation (AAAI 2024 · similarity 0.86)
- Beyond Background Shift: Rethinking Instance Replay in Continual Semantic Segmentation (CVPR 2025 · similarity 0.84)
- Recall-Oriented Continual Learning with Generative Adversarial Meta-Model (AAAI 2024 · similarity 0.84)
- ASAM: Boosting Segment Anything Model with Adversarial Tuning (CVPR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.