On-device training enables the model to adapt to user-specific data by fine-tuning a pre-trained model locally. As embedded devices become ubiquitous, on-device training is increasingly essential since users can benefit from the personalized model without transmitting data and model parameters to the server. Despite significant efforts toward efficient training, ondevice training still faces a major challenge: The prohibitive cost of multi-layer backpropagation strains the limited resources of tiny devices. In this paper, we propose an algorithm-system cooptimization framework TinyMP that enables self-adaptive on-tiny-device model personalization. To mitigate backpropagation costs, we introduce Gradient Condensing to condense the gradient map structure, significantly reducing the computational complexity and memory consumption of backpropagation while preserving model performance. To further reduce computation overhead, we propose Alternant Partial Update, a mechanism that locally and alternatively selects essential parameters to update without requiring retraining or offline evolutionary search. Our framework is evaluated through extensive experiments using various CNN models (e.g., MobileNetV2, MCUNet) on embedded devices with minimal resources (e.g., OpenMV-H7 with less than 1MB SRAM and 2 MB Flash). Experimental results show that our framework achieves up to $2.4 \times$ speedup, 80.8% memory saving, and 30.3% accuracy improvement on downstream tasks, outperforming SOTA approaches.

Acoustic scene classification under lowcomplexity conditions presents unique challenges due to limited data availability and computational constraints. To address this, we propose a lightweight and effective ASC framework that integrates multi-scale feature fusion with attention mechanisms, specifically designed for resource-constrained environments. The model begins by extracting time-frequency representations using 3D convolutional layers, which efficiently capture spatial and temporal patterns in audio signals. To further emphasize critical information, an attention mechanism guided by timesemantic neighbors dynamically assigns three-dimensional weights to salient features, highlighting the most informative components in the time-frequency space. In addition, a pyramid pooling module with multi-scale dilated convolutions is employed to aggregate both local and global contextual cues, enhancing the model's ability to detect distinctive acoustic events at specific frequency ranges. This hierarchical design improves feature expressiveness without introducing significant computational overhead. Extensive experiments on multiple benchmark datasets demonstrate that our method not only achieves competitive or superior classification accuracy compared to state-of-the-art techniques, but also maintains low model complexity and inference cost, making it highly suitable for deployment in real-world, low-complexity scenarios.

Reversible data hiding in encrypted images (RDHEI) is essential for safeguarding sensitive information within the encrypted domain. In this study, we propose an intelligent pixel predictor based on a residual group block and a spatial attention module, showing superior pixel prediction performance compared to existing predictors. Additionally, we introduce an adaptive joint coding method that leverages bit-plane characteristics and intra-block pixel correlations to maximize embedding space, outperforming single coding approaches. The image owner employs the presented intelligent predictor to forecast the original image, followed by encryption through additive secret sharing before conveying the encrypted image to data hiders. Subsequently, data hiders encrypt secret data and embed them within the encrypted image before transmitting the image to the receiver. The receiver can extract secret data and recover the original image losslessly, with the processes of data extraction and image recovery being separable. Our innovative approach combines an intelligent predictor with additive secret sharing, achieving reversible data embedding and extraction while ensuring security and lossless recovery. Experimental results demonstrate that the predictor performs well and has a substantial embedding capacity. For the Lena image, the number of prediction errors within the range of [−5, 5] is as high as 242 500 and our predictor achieves an embedding capacity of 4.39 bpp.