Selective encryption hides only specified categories in UAV images
The claim
The framework encrypts only user-specified sensitive categories while leaving other scene content unchanged
Published as: “A Category-Selective Encryption Framework for Privacy Protection of UAV Images”
So what
Enables UAV image transmission that protects privacy of selected objects while keeping the rest of the scene usable for downstream analytics, suitable for real-time surveillance deployments.
The five-minute version
The framework encrypts only user-specified categories in UAV images using YOLO11 detection and two interchangeable schemes-adaptive pixelation (TPE) with learnable blending and a XOR stream cipher. XOR encryption yields lossless reconstruction (SSIM = 1, infinite PSNR) while TPE achieves high but slightly lower quality (SSIM ≈ 0.979, PSNR ≈ 38.5 dB for pedestrians). Experiments on the public VisDrone dataset show the selective approach preserves background utility and runs in real-time (≈0.5 ms for XOR, ≈1 ms encryption for TPE), though detection accuracy is modest (≈27.6%/15.5%) and TPE decryption incurs ≈30 ms latency.
What they found 3
- XOR encryption provides lossless reconstruction of encrypted regions.
“XOR encryption achieves an SSIM of 1 and infinite PSNR under both encrypted region settings, which demonstrates that XOR enables lossless pixel-level reconstruction without any image distortion after decryption.”
- TPE decryption quality degrades when more categories are encrypted.
“The decryption quality remains favorable when only pedestrians are encrypted, while the restoration performance degrades when both pedestrians and vehicles are encrypted due to the expanded encrypted area.”
- XOR encryption and decryption run in about half a millisecond.
“XOR pipeline is a handcrafted lightweight bitwise operation without neural network components, achieving only 0.5 ms latency for both encryption and decryption under the unified experimental environment.”
What undercuts it the paper’s own results
- The underlying detector attains only 27.6% and 15.5% accuracy on VisDrone, limiting reliable identification of all sensitive objects.
- TPE decryption requires a deep network (50.29 M parameters) and incurs ~30 ms latency, which may be too slow for strict real-time constraints.
Stated limits 3
- Quantitative reliability tests on the VisDrone test set [19] show that the pre-trained model achieves = 27.6% and = 15.5%, which is standard for lightweight detectors in complex UAV scenes with dense small objects and varying altitudes.
- Future work will further optimize the network inference speed and encryption security strength, expand the applicable scope of the algorithm, and adopt more robust detection and joint detection-encryption optimization strategies to mitigate privacy leakage risks caused by undetected small, occluded or poorly illuminated sensitive objects.
- In contrast, TPE decryption employs a deep convolutional reconstruction network with 50.29 million trainable parameters to recover distorted texture details, incurring substantial matrix computation and memory access overhead and a latency of around 30 ms.
What’s new here 4
- Semantic-level selective encryption driven by object detection rather than class-agnostic saliency.
- Dual, switchable encryption modes: learnable pixelation (TPE) and exact XOR stream cipher.
- Joint training of TPE and an enhanced U-Net with a composite loss that balances privacy strength and restoration quality.
- Real-time processing with lightweight XOR encryption and targeted computation only on detected sensitive regions.
Artifacts & methods 5
- YOLO11n object detectorsystem
- Thumbnail-Preserving Encryption (TPE) modulesystem
- XOR stream cipher encryption modulesystem
- Enhanced U-Net restoration networksystem
- VisDrone2019 datasetdataset