DARDA: Domain-Aware Real-Time Dynamic Neural Network Adaptation

Shahriar Rifat, Jonathan D. Ashdown, Francesco Restuccia

Introduction

Deploying Deep Neural Networks (DNNs) in real-world edge environments presents significant challenges. Unlike controlled laboratory settings, field deployments often encounter environmental "corruptions"—such as fog, glare, or sensor noise—that cause a distribution shift. This shift occurs when runtime data deviates from the training distribution, leading to degraded model performance. While Test-Time Adaptation (TTA) aims to address this by updating models on-the-fly using unlabeled data, existing methods often suffer from high computational costs, error accumulation (catastrophic forgetting), and poor performance on non-IID (Independent and Identically Distributed) data streams typical of real-world video feeds.

The DARDA Framework

To overcome these hurdles, we present DARDA (Domain-Aware Real-Time Dynamic Neural Network Adaptation). DARDA shifts the paradigm from reactive fine-tuning to proactive domain management. Instead of continuous, resource-intensive model updates, DARDA identifies specific environmental corruptions and dynamically switches the network to a pre-optimized "sub-network" state.

Overview of the DARDA framework architecture, illustrating the flow from corruption extraction to dynamic sub-network adaptation.

Key Methodological Components

  • Corruption Extractor: An unsupervised module designed to decouple environmental noise from image content. By analyzing the correlation between adjacent pixels, it isolates corruption signatures without requiring clean reference images.
  • Cross-Modal Latent Space: Extracted corruption signatures and sub-network "fingerprints" (generated using fixed Gaussian noise) are mapped into a shared latent space.
  • Dynamic Switching: At runtime, the system identifies the current corruption signature and activates the closest corresponding sub-network in the latent space.
  • Corruption-Aware Memory Bank: A small, label-balanced buffer maintains recent samples to refine Batch Normalization (BN) statistics, ensuring stability without the need for full backpropagation on every frame.

Experimental Evaluation

DARDA was evaluated using CIFAR-10, CIFAR-100, and TinyImagenet datasets, augmented with 15 known and 4 unseen corruption types (e.g., snow, frost, JPEG artifacts). Testing was conducted on edge hardware, specifically the NVIDIA Jetson Nano and Raspberry Pi 5.

Key Results and Discussion

DARDA demonstrates superior performance and efficiency compared to state-of-the-art TTA methods like RoTTA.

Resource consumption increase relative to inference-only execution (left) and adaptation latency across different corruption types for Jetson Nano and Raspberry Pi (right).
  • Accuracy: DARDA achieved significant accuracy improvements: 10.4% on CIFAR-10, 5.7% on CIFAR-100, and 4.4% on TinyImagenet.
  • Resource Efficiency: The framework reduced energy consumption by 1.74× and average cache memory footprint by 2.64× compared to previous methods.
  • Robustness: DARDA maintains high performance even with highly correlated non-IID data, a critical requirement for robotics and UAV applications where temporal correlation is high.

Conclusion

DARDA provides a robust and energy-efficient solution for deploying AI at the edge. By proactively mapping environmental domains to specialized sub-networks, it avoids the pitfalls of continuous adaptation. Future research will explore extending this framework to more complex vision tasks, such as object detection and semantic segmentation, to further enhance its utility in diverse real-world scenarios.

Summary updated

Abstract

Test Time Adaptation (TTA) has emerged as a practical solution to mitigate the performance degradation of Deep Neural Networks (DNNs) in the presence of corruption/ noise affecting inputs. Existing approaches in TTA continuously adapt the DNN, leading to excessive resource consumption and performance degradation due to accumulation of error stemming from lack of supervision. In this work, we propose Domain-Aware Real-Time Dynamic Adaptation (DARDA) to address such issues. Our key approach is to proactively learn latent representations of some corruption types, each one associated with a sub-network state tailored to correctly classify inputs affected by that corruption. After deployment, DARDA adapts the DNN to previously unseen corruptions in an unsupervised fashion by (i) estimating the latent representation of the ongoing corruption; (ii) selecting the sub-network whose associated corruption is the closest in the latent space to the ongoing corruption; and (iii) adapting DNN state, so that its representation matches the ongoing corruption. This way, DARDA is more resource efficient and can swiftly adapt to new distributions caused by different corruptions without requiring a large variety of input data. Through experiments with two popular mobile edge devices - Raspberry Pi and NVIDIA Jetson Nano - we show that DARDA reduces energy consumption and average cache memory footprint respectively by 1.74x and 2.64x with respect to the state of the art, while increasing the performance by 10.4%, 5.7% and 4.4% on CIFAR-10, CIFAR-100 and TinyImagenet.
BibTeX
@inproceedings{rifat2025darda,
  author = {Shahriar Rifat and Jonathan D. Ashdown and Francesco Restuccia},
  title = {{DARDA: Domain-Aware Real-Time Dynamic Neural Network Adaptation}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2025},
  doi = {10.1109/wacv61041.2025.00194},
}
Figure