Catch-Only-One: Non-Transferable Examples for Model-Specific Authorization
Organizations: The University of Queensland · CSIRO · Monash University · City University of Hong Kong
Abstract
Recent AI regulations increasingly emphasize the need for mechanisms that preserve the utility of data for AI innovation while preventing misuse, particularly by enforcing purpose limitation in downstream AI applications. In practice, enforcing this principle remains challenging, as released data can be trivially fed into arbitrary models beyond its declared intent. Existing approaches attempt to mitigate this risk by either perturbing data or retraining models to limit unintended use. These strategies, however, offer no protection against inference by unknown or externally trained models, or fundamentally rely on control over the training or deployment. In this work, we introduce non-transferable examples (NTEs), recoded data that act as a task-level "ciphertext" decodable only by a designated model. Whereas adversarial examples exploit directions of high model sensitivity, NTEs leverage the complementary insensitive subspace. We propose a training-free, data-agnostic method that recodes data within a model-specific low-sensitivity subspace, preserving outputs for the authorized model while degrading unauthorized ones through subspace misalignment. We establish formal bounds certifying authorized-model fidelity and showing that unauthorized degradation scales with measurable spectral misalignment between models. Empirically, NTEs preserve performance across diverse vision backbones and state-of-the-art vision-language models under common preprocessing, while unauthorized models collapse even under adaptive reconstruction attacks. These results establish NTEs as a practical means to preserve intended data utility while preventing unauthorized exploitation. Our project is available at https://trusted-system-lab.github.io/model-specificity.
Figures & tables
| CIFAR-10 | ImageNet | |||||||||
| ResNet-50 | ViT-B | SwinV2-T | DeiT-B | MambaVision-T | ResNet-50 | ViT-B | SwinV2-T | DeiT-B | MambaVision-T | |
| Baseline | 98.2% | 98.8% | 96.1% | 96.1% | 96.9% | 80.3% | 81.6% | 80.9% | 79.9% | 82.4% |
| ResNet-50 | 97.7 % | 12.5% | 13.6% | 9.4% | 10.0% | 80.2 % | 0.0% | 0.1% | 0.1% | 0.1% |
| ViT-B | 10.5% | 98.7 % | 9.3% | 12.1% | 9.7% | 0.0% | 81.3 % | 0.0% | 0.0% | 0.0% |
| SwinV2-T | 15.6% | 11.7% | 88.4 % | 20.6% | 18.8% | 9.1% | 4.3% | 71.7 % | 7.4% | 12.9% |
| DeiT-B | 9.4% | 9.8% | 9.8% | 96.1 % | 5.5% | 0.0% | 0.0% | 0.0% | 79.3 % | 1.0% |
| CIFAR-10 | ImageNet | |||||||||
| Plain | DP | FHE 3 | AlgoSpec | Nte (Ours) | Plain | DP | FHE 3 | AlgoSpec | Nte (Ours) | |
| ResNet-50 | 98.2% | 59.8% 1 | 87.8% 3 | 6.4% | 97.7% | 80.3% | 63.1% 1 | – | 0.1% | 80.2% |
| ViT-B | 98.8% | – 1 | – | 10.0% | 98.7% | 81.6% | – 1 | – | 0.0% | 81.3% |
| Protection | ✗ | ✗ 2 | ✓ | ✗ 2 | ✓ | ✗ | ✗ 2 | ✓ | ✗ 2 | ✓ |
| InternVL3-1B (authorized) | Qwen2.5-VL-3B (unauthorized) | |||||||||||||
| Overall | AR | CP | FP-C | FP-S | LR | RR | Overall | AR | CP | FP-C | FP-S | LR | RR | |
| Baseline | 72.7% | 77.4% | 82.4% | 58.7% | 79.4% | 50.3% | 66.8% | 78.8% | 81.3% | 83.1% | 69.2% | 84.9% | 66.5% | 75.4% |
| Nte (Ours) | 72.6% | 77.8% | 82.0% | 58.3% | 78.9% | 50.9% | 67.3% | 18.3% | 29.2% | 15.6% | 17.0% | 12.8% | 17.3% | 22.3% |
| Attacker Model | ResNet-50 | ViT-B | SwinV2-T | DeiT-B | MambaVision-T |
| Clean | 98.2% | 98.8% | 96.1% | 96.1% | 96.9% |
| Attacker | 11.1% | 10.9% | 10.5% | 13.8% | 10.9% |
| Nte -Attacker | 10.3% | 10.7% | 10.3% | 10.9% | 11.2% |
| Model | SVD (ms) 1 |
| ResNet-50 | |
| ViT-B | |
| SwinV2-T | |
| DeiT-B | |
| MambaVision-T | |
| Encoding / image 2 | 0.11 ms |
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
| Max | Min | Mean | Median | |
| ResNet-50 | 3.8203 | 7.77 10 8 | 1.0242 | 0.5904 |
| ViT-B | 9.5686 | 3.68 10 5 | 0.5666 | 0.0482 |
| PSNR (dB) | L0 | L1 | L2 | L3 | L4 | L5 | L6 |
| 40 | 0.0032 | 0.0097 | 0.0014 | 0.0062 | 0.0043 | 0.0210 | 0.0008 |
| 35 | 0.0035 | 0.0107 | 0.0017 | 0.0072 | 0.0050 | 0.0247 | 0.0010 |
| 30 | 0.0071 | 0.0200 | 0.0031 | 0.0146 | 0.0102 | 0.0498 | 0.0023 |
| 25 | 0.0134 | 0.0331 | 0.0052 | 0.0272 | 0.0184 | 0.0859 | 0.0040 |
| 20 | 0.0112 | 0.0367 | 0.0068 | 0.0282 | 0.0207 | 0.1016 | 0.0046 |
| 15 | 0.0240 | 0.0680 | 0.0104 | 0.0499 | 0.0349 | 0.1672 | 0.0079 |
| CIFAR-10 | ImageNet | ||||||||||||
| ResNet-50 | ViT-B | SwinV2-T | DeiT-B | MambaVision-T | ResNet-50 | ViT-B | SwinV2-T | DeiT-B | MambaVision-T | ||||
| ResNet-50 | ImageNet | 13.3% | 9.4% | 8.2% | 9.4% | 7.8% | CIFAR-10 | 1.2% | 0.0% | 4.4% | 0.0% | 0.0% | |
| ViT-B | 10.1% | 9.6% | 10.1% | 11.5% | 9.8% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | |||
| SwinV2-T | 10.5% | 10.2% | 21.0% | 11.3% | 14.5% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | |||
| DeiT-B | 12.5% | 9.3% | 7.0% | 14.8% | 10.9% | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | |||
| MambaVision-T | 14.5% | 5.1% | 12.5% | 9.8% | 8.2% | 0.0% | 0.0% | 1.9% | 0.0% | 0.0% | |||
| BERT-base | RoBERTa-base | |||||||||||||
| CoLA | MNLI | QNLI | QQP | RTE | SST-2 | STS-B | CoLA | MNLI | QNLI | QQP | RTE | SST-2 | STS-B | |
| Baseline | 54.2 | 83.4 | 90.5 | 90.1 | 60.3 | 91.6 | 87.1 | 53.8 | 87.7 | 92.8 | 90.9 | 66.1 | 94.5 | 87.5 |
| Nte (Ours) | 54.5 | 82.9 | 89.5 | 89.6 | 60.3 | 89.6 | 86.9 | 55.5 | 87.5 | 92.4 | 90.8 | 65.3 | 94.6 | 87.2 |
| Unauthorized | 33.4 | 31.8 | 50.5 | 36.8 | 47.3 | 50.9 | 60.2 | 37.1 | 35.4 | 49.5 | 63.2 | 52.7 | 49.1 | 63.5 |
| Prompt Please describe the image in detail. | ||
| (a) Nte input | (b) clean input | |
| InternVL3 authorized | In the image, a person is sitting on a sandy beach, smiling and playing with a dog. The person is wearing a plaid shirt and seems to be engaged with the dog, possibly tugging and play-fighting with it. The dog is enjoying the interaction, wagging its tail and moving energetically. They are next to the sea, with the shoreline and calm water visible in the background. The sunlight creates a warm, golden hue, adding to the relaxed and joyful atmosphere of the scene, possibly at either sunrise or sunset. | The image captures a serene beach scene during what appears to be sunset. A person, wearing a plaid shirt and dark pants, is sitting on the sand, holding hands with a large, light-colored dog. The dog is attentively playing with the person, reaching out its paw with joy and companionship… |
| Qwen2.5-VL unauthorized | The image appears to be a pixelated or low-resolution representation of a square object. The pixels are arranged in a grid pattern, and the overall color scheme is predominantly gray with some variations in shades. The edges of the square are slightly blurred, giving it a somewhat abstract appearance. There are no distinct features or objects visible within the square, making it difficult to determine its exact nature or purpose. | The image depicts a serene beach scene with a person and a dog. The person is sitting on the sandy beach, facing the ocean. They are wearing a plaid shirt and dark pants, and they have long hair. The dog, which appears to be a Labrador Retriever, is sitting on the sand and is interacting … |
| Prompt Please describe this image in a single concise paragraph. | ||
| (a) Nte input | (b) clean input | |
| InternVL3 authorized | This image shows a super busy city street packed with traffic and people, with lots of yellow taxis in the foreground along with other cars all stuck in congestion. The street is lined with tall buildings on both sides, creating a canyon effect, and you can see many traffic lights, street signs, and American flags hanging from poles. There are pedestrians on the sidewalks, especially on the left side of the image, and farther down the street the traffic continues into the distance, getting hazier because of the depth and maybe a bit of atmospheric blur. Overall, it looks like a classic dense downtown scene in a big U.S. city, with heavy traffic, high-rises, and a lot of visual clutter from signs, ads, and lights. | The image depicts a bustling urban street scene, likely in a major city known for its iconic yellow taxis. The street is filled with numerous vehicles, including cars and taxis, indicating heavy traffic. The buildings lining the street are tall and modern, suggesting a downtown area. Street signs and traffic lights are visible, directing the flow of traffic. The sky appears overcast, adding to the busy atmosphere. The overall scene conveys a sense of congestion and the dynamic nature of city life. |
| GPT-5 unauthorized | The image is a densely packed, full-frame pattern made up of extremely small, repeating colorful shapes that create a fine, almost pixel-like texture; from a distance it appears as multicolored visual static with a subtle grid structure, dominated by greens, purples, yellows, and blues distributed uniformly across the entire square. | The image shows a busy city avenue lined with tall buildings, dense traffic, and numerous yellow taxis, with traffic lights, street signs, flags, and storefronts creating a lively, crowded urban scene that fades into a hazy skyline in the distance. |
| Kimi k1.5 | The image displays a patterned texture resembling a fabric or material with a consistent, tightly woven design. The pattern is composed of small, interlocking shapes that create a complex, geometric lattice. The color palette is muted, featuring a blend of grays, greens, and subtle hints of other colors, giving it an almost camouflage-like appearance. The texture appears flat and uniform throughout, suggesting a synthetic or manufactured origin rather than a natural one. | The image captures a busy New York City street filled with yellow taxis and various vehicles, surrounded by tall buildings adorned with signs and American flags. Traffic lights and street signs, including directions and regulations, are visible, contributing to the organized chaos typical of a bustling urban environment. The overcast sky softens the scene, highlighting the city’s dynamic nature. |
| Sonnet 4.5 | This image appears to be random digital noise or static, displaying a uniform pattern of multicolored pixels (red, green, blue, and gray) distributed across the entire frame with no discernible subject, pattern, or structure. It resembles the visual “snow” or grain you might see on an old television with no signal, or could be the result of a corrupted image file, sensor noise from a camera, or an intentionally generated random noise pattern. | This is a vibrant street photograph of a busy Manhattan avenue in New York City, showcasing the iconic urban canyon effect with tall buildings lining both sides of the street. The scene is packed with yellow taxi cabs and other vehicles in dense gridlock, with multiple traffic lights hanging overhead on wires. The atmospheric haze in the distance adds depth to the composition, while pedestrians and commercial signage contribute to the bustling energy. The photo perfectly captures that quintessential NYC feeling - crowded, vertical, and pulsing with constant movement and organized chaos. |
| Grok 4 | This image is a stereogram (autostereogram), a 2D pattern of repeating green, purple, brown, and yellow pixel-like dots designed to create a 3D optical illusion when viewed by crossing or diverging the eyes, revealing a hidden image embedded within the structured, tiled texture. | This image captures a bustling urban street scene in a major city, likely New York City, characterized by heavy traffic and iconic yellow taxis. The street is flanked by tall, modern buildings, some of which feature prominent signage and advertisements. Pedestrians can be seen on the sidewalks, and the overall atmosphere is one of a dense, busy metropolitan environment. The lighting suggests it might be daytime, but with an overcast or slightly hazy sky. |
| Method | Input | SR-ResNet 1 | SR-ResNet 2 |
| Gaussian | 15.8 | 33.5 | 35.9 |
| Nte s | 10.2 | 10.5 | 10.9 |
| Nte s 3 | 10.7 | 10.8 | 11.1 |
| Training length (ViT-B, full data) | 1 epoch | 3 epochs | 10 epochs | 30 epochs |
| Accuracy | 11.1% | 10.7% | 10.1% | 10.7% |
| Released data (ViT-B, 30 epochs) | 10% | 25% | 50% | 100% |
| Accuracy | 10.1% | 10.7% | 10.3% | 10.7% |
Privacy Protection in VLMs
Privacy protection in vision-language models limits sensitive information exposure while preserving useful image understanding and multimodal reasoning.
See all 43 papers