A Category-Selective Encryption Framework for Privacy Protection of UAV Images

I. INTRODUCTION

• UAV images may expose faces, pedestrians, and license plates if leaked during wireless transmission or cloud storage.Deep-learning algorithms can also identify human contours and behavioral activities in aerial images.Source: §I · p. 1
• Existing UAV encryption schemes generally use a fixed mode, while saliency-based methods obscure salient regions and can remove information needed for subsequent analysis.These approaches do not support discriminative encryption of semantic targets such as pedestrians and vehicles.Source: §I · p. 1
• The framework combines multi-category detection with two switchable encryption strategies in a closed-loop pipeline for selective encryption and restoration.It aims to preserve scene information under privacy guarantees.Source: §I · pp. 1–2
• YOLO11 detects all 80 target classes, while encryption is applied only to user-defined sensitive categories.The framework also uses collaborative optimization across detection, encryption, and restoration with a composite loss to improve TPE decryption quality.Source: §I · p. 2

II. RELATED WORKS

• Existing saliency-driven privacy schemes, region-selective encryption, and basic anonymization do not provide category-customized encryption with complete high-fidelity recovery pipelines.The section introduces TPE and XOR as complementary encryption technologies.Source: §II · p. 2

A. TPE

• TPE preserves an image thumbnail by encrypting sub-blocks while retaining their pixel averages, with block size controlling the security–usability trade-off.Prior methods include exact and approximate TPE; TPE-GAN improves ciphertext usability, though reconstruction quality remains open to improvement.Source: A. TPE · p. 2

B. XOR Stream Cipher Fundamentals

• XOR stream ciphers can restore the original image losslessly using the identical seed key, but traditional versions lack semantic target screening and obscure the overall image.Their low computational overhead suits lightweight terminal devices.Source: B. XOR Stream Cipher Fundamentals · p. 2

III. PROPOSED SCHEME

• The framework forms a closed-loop pipeline with YOLO11 detection, TPE and XOR encryption modules, and corresponding decryption and restoration modules.The flowchart depicts the overall proposed scheme.Source: §III · p. 2, Figure 1 · p. 3

A. Object Detection Module

• YOLO11n, fine-tuned on VisDrone2019, detects and categorizes aerial-scene objects; users select sensitive categories at runtime for encryption.Filtered bounding boxes are normalized for subsequent mask generation.Source: A. Object Detection Module · p. 2

B. Selective Encryption Schemes

• TPE and XOR share an input interface and modify only detected sensitive regions, leaving non-sensitive background pixels identical to the original.A multi-stage joint-training strategy is designed only for the TPE branch.Source: §III · p. 2, B. Selective Encryption Schemes · p. 3

1) TPE Encryption:

• TPE applies pixelation to sensitive target regions and relies on a trained Enhanced U-Net decoder to restore visual content.The learnable blending coefficient balances privacy protection strength and restoration quality during joint training.Source: 1) TPE Encryption · p. 3
• Larger targets receive larger mosaic blocks, while distant small targets use finer blocks to maintain consistent visual encryption effects across object scales.Block size is calculated dynamically from target-box width and height, with kmin = 6 and scaling factor d = 8.Source: 1) TPE Encryption · p. 3
• Average pooling reduces each k × k pixel block to one average color, then nearest-neighbor upsampling restores the patch’s spatial dimensions to create mosaic texture.The pooling kernel and stride are both k.Source: 1) TPE Encryption · p. 3
• The encrypted patch linearly blends pixelated and original patches using α, constrained to [0, 1].At α = 1 the patch is fully pixelated; at α = 0 it remains identical to the original.Source: 1) TPE Encryption · p. 3

• The user key is hashed with SHA-256 to seed PCG-64, which generates a three-channel noise map for the input image.The complete noise map is generated at once, ensuring consistent results when target boxes overlap.Source: 2) XOR Stream Cipher Encryption · p. 3
• Applying bitwise XOR between the masked image pixels and corresponding noise pixels encrypts the sensitive regions.The XOR operation is involutive, enabling mathematically lossless reversibility.Source: 2) XOR Stream Cipher Encryption · p. 3
• XOR encryption modifies only pixels marked as sensitive by the binary mask, leaving background pixel values unchanged.Source: 2) XOR Stream Cipher Encryption · p. 3

C. Decryption and Restoration Module

• The decryption and restoration module supplies branch-specific recovery pipelines for images encrypted with TPE or XOR.Source: C. Decryption and Restoration Module · p. 3

1) TPE Decryption:

• TPE restoration uses an Enhanced U-Net to recover texture from encrypted images because learnable mosaic blur has no straightforward mathematical inverse.Multi-scale skip connections suppress mosaic artifacts and restore target details without changing background pixels; the decoder and TPE module are jointly trained.Source: 1) TPE Decryption · p. 3

2) XOR Stream Cipher Decryption:

• XOR decryption regenerates pseudo-random noise from the original secret key and applies XOR again within the encryption mask to recover lossless uint8 pixels.This path requires no neural network computation.Source: 2) XOR Stream Cipher Decryption · p. 3

3) Post-processing Coordinate Inversion:

• After either restoration branch, cached LetterBox metadata is used to remove symmetric padding and rescale images to their original resolution.This inverse transformation addresses deformation introduced during preprocessing.Source: 3) Post-processing Coordinate Inversion · p. 4

D. Joint Training Strategy

• Joint training performs gradient optimization only on the TPE encryption branch to balance detection accuracy, privacy protection, and restoration quality.Source: D. Joint Training Strategy · p. 4

1) Parameter Freezing and Update Rules:

• The YOLO11n detector remains frozen; training updates only TPE’s blending coefficient α and Enhanced U-Net weights, with masked sensitive pixels alone contributing to loss and gradients.Excluding background pixels from optimization preserves consistency with the original image and reduces trainable parameters and GPU memory use.Source: 1) Parameter Freezing and Update Rules · p. 4

2) Loss Function Design:

• The total loss combines reconstruction loss with weighted encryption regularization, and the regularization weight λα is set to 0.6.Reconstruction uses L1 error only within masked sensitive regions, while encryption regularization pushes α toward 1 to prevent excessive weakening of the mosaic.Source: 2) Loss Function Design · p. 4

• Experiments are conducted on VisDrone, using a YOLO11n detector fine-tuned on that dataset and held fixed during joint training.Source: §IV · p. 4
• Fig. 2 compares pedestrian-only encryption with pedestrian-and-vehicle encryption; within each group, TPE is shown above XOR.Source: Figure 2 · p. 4
• The TPE setup uses minimum block size kmin = 6 and scaling factor d = 8, while Enhanced U-Net uses SE channel attention with compression ratio r = 8.Training uses AdamW, learning rate 1 × 10−4, batch size 8, λα = 0.6, and up to 120 epochs with early stopping.Source: §IV · p. 4

A. Visual Effect Analysis

• Both schemes encrypt only the selected sensitive regions and leave the background unchanged; TPE obscures targets with mosaic blur, while XOR uses pseudo-random noise.Enhanced U-Net restoration reconstructs encrypted target textures and contours; XOR permits strict pixel-level reconstruction with the correct key.Source: A. Visual Effect Analysis · p. 4

B. Image Quality Analysis

• XOR achieves SSIM of 1 and infinite PSNR under both encrypted-region settings, while TPE reconstruction degrades as the encrypted area expands.The comparison uses PSNR and SSIM on the VisDrone test set.Source: B. Image Quality Analysis · pp. 4–5, Table I · p. 5

C. Comparative Analysis

• Prior saliency-driven, region-selective, and anonymization approaches are described as lacking designated-category customization and complete high-fidelity recovery pipelines.Source: C. Comparative Analysis · p. 5
• HF-TPE and TPE-GAN encrypt full images, distorting roads and buildings, whereas the proposed framework limits encryption to predefined sensitive objects.The cited comparison associates background distortion with impaired subsequent scene recognition and substantial computational overhead.Source: C. Comparative Analysis · p. 5
• The proposed XOR branch achieves SSIM of 1 and infinite PSNR, and both proposed branches outperform their corresponding baselines in reconstruction quality.The comparison reports PSNR and SSIM across encryption schemes and encrypted regions.Source: Table II · p. 5, C. Comparative Analysis · p. 5

D. Detection Performance and Privacy Implications

• On the VisDrone test set, the pre-trained model achieves mAP@0.5 = 27.6% and mAP@0.5:0.95 = 15.5%; distant tiny targets have limited recognizable details, yet privacy leakage may still occur through cross-frame contextual reasoning.Source: D. Detection Performance and Privacy… · p. 5

E. Efficiency Analysis

• XOR encryption and decryption each take 0.5 ms; TPE encryption takes approximately 1 ms, while its 50.29-million-parameter decoder takes around 30 ms.Both modules restrict computation to YOLO11n sensitive bounding boxes, reducing the number of processed pixels.Source: Table III · p. 6, E. Efficiency Analysis · p. 6

V. CONCLUSION

• The framework encrypts designated categories and preserves basic image content; its conclusion identifies reduced redundant computation and suitability for real-time UAV image transmission.Future work targets inference speed, encryption security, broader applicability, and more robust detection and joint detection-encryption optimization.Source: §V · p. 6