Research map / Label-efficient and robust learning / Domain adaptation and generalization
Your Transferability Barrier is Fragile: Free-Lunch for Transferring the Non-Transferable Learning
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.
Shows that non-transferable learning (NTL) models are impaired on third-party domains and proposes TransNTL, an impairment-repair self-distillation plus sharpness-aware attack that recovers target-domain accuracy using 10% of source data, along with a defense.
Problem
Non-transferable learning (NTL) restricts a model's generalization to a target domain for IP protection, but the robustness of this "transferability barrier" against an attacker with few source-domain data has not been studied.
Approach
Perturbed source-domain images act as third-party domains that reproduce the over-confident, target-class-consistent impairment patterns of NTL models; fine-tuning with a self-distillation loss on these perturbations, plus a sharpness-aware term, repairs the impairment and breaks the barrier.
Key results
- TransNTL recovers target-domain accuracy with 10% source data, increasing it by up to 49.2% for NTL and 55.0% for CUTI in the main transfer experiments.
- On NTL-based ownership verification, TransNTL raises patched-domain accuracy by up to 71.7%, and on applicability authorization by up to 28.7%.
- The proposed defense (R-CUTI) limits the post-attack target accuracy gain to 0.0% on CIFAR10->STL10 and 16.6% on VisDA-T->VisDA-V, versus 52.4% and 20.9% for CUTI.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply TransNTL fine-tuning with 10% source data | NTL and CUTI · MNIST→Digits | target accuracy: +25.4; +39.7 | Helped |
| Apply TransNTL fine-tuning with 10% source data | NTL and CUTI · SVHN→Digits | target accuracy: +37.6; +44.4 | Helped |
| Apply TransNTL fine-tuning with 10% source data | NTL and CUTI · CIFAR10→STL10 | target accuracy: +39.6; +52.4 | Helped |
| Apply TransNTL fine-tuning with 10% source data | NTL and CUTI · STL10→CIFAR10 | target accuracy: +49.2; +55.0 | Helped |
| Apply TransNTL fine-tuning with 10% source data | NTL and CUTI · VisDA-T→VisDA-V | target accuracy: +17.0; +20.9 | Helped |
| Add self-distillation loss to vanilla fine-tuning (ablation) | TransNTL on NTL and CUTI · CIFAR10→STL10 | target accuracy: +34.7; +49.4 | Helped |
| Add self-distillation loss to vanilla fine-tuning (ablation) | TransNTL on NTL and CUTI · VisDA-T→VisDA-V | target accuracy: +11.4; +18.0 | Helped |
| Add sharpness term to fine-tuning with self-distillation (ablation) | TransNTL on NTL and CUTI · CIFAR10→STL10 | target accuracy | Helped |
| Add sharpness term to fine-tuning with self-distillation (ablation) | TransNTL on NTL and CUTI · VisDA-T→VisDA-V | target accuracy | Helped |
| Use vanilla fine-tuning alone, without self-distillation or sharpness (ablation) | TransNTL on NTL and CUTI · CIFAR10→STL10 and VisDA-T→VisDA-V | target accuracy: 0.0; +0.1; +0.0; +0.1 | No clear effect |
Surprising findings
- NTL models make over-confident predictions and collapse third-party domain predictions onto the implicit target-domain class, even under slight distribution shift from the source domain.
- Simple perturbed source-domain data can recover target-domain performance without any target-domain knowledge.
Limitations
- Evaluated only on small image classification benchmarks (digits, CIFAR10/STL10, VisDA).
- Source-domain accuracy drops for some settings (e.g., 6.7% for NTL on VisDA-T) under self-distillation alone.
- Defense is evaluated only against CUTI on two tasks.
Experimental setup
Datasets: MNIST MNIST-M SVHN SYN-D CIFAR10 STL10 VisDA-2017
Benchmarks: MNIST->Digits SVHN->Digits CIFAR10->STL10 STL10->CIFAR10 VisDA-T->VisDA-V
Metrics: top-1 accuracy on source domain top-1 accuracy on target domain
Baselines: FTAL RTAL FP NAD i-BAU FT-SAM NTL CUTI
Builds on: non-transferable learning CUTI sharpness-aware minimization adaptive SAM
Closest papers by meaning
- Jailbreaking the Non-Transferable Barrier via Test-Time Data Disguising (CVPR 2025 · similarity 0.91)
- Non-Transferable Learning: A New Approach for Model Ownership Verification and Applicability Authorization (ICLR 2022 · similarity 0.86)
- Improving Non-Transferable Representation Learning by Harnessing Content and Style (ICLR 2024 · similarity 0.85)
- Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.84)
- When Data-Free Knowledge Distillation Meets Non-Transferable Teacher: Escaping Out-of-Distribution Trap is All You Need (ICML 2025 · similarity 0.82)
- Catch-Only-One: Non-Transferable Examples for Model-Specific Authorization (NeurIPS 2026 · similarity 0.82)
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.