Split Computing enables collaborative inference between edge devices and the cloud by partitioning a deep neural network into an edge-side head and a server-side tail, reducing latency and limiting exposure of raw input data. However, inference performance often degrades in practical deployments due to user-specific data distribution shifts, unreliable communication, and privacy-oriented perturbations, especially in closed environments where model architectures and parameters are inaccessible. To address this challenge, we propose SALT (Split-Adaptive Lightweight Tuning), a lightweight adaptation framework for closed Split Computing systems. SALT introduces a compact client-side adapter that refines intermediate representations produced by a frozen head network, enabling effective model adaptation without modifying the head or tail networks or increasing communication overhead. By modifying only the training conditions, SALT supports multiple adaptation objectives, including user personalization, communication robustness, and privacy-aware inference. Experiments using ResNet-18 on CIFAR-10 and CIFAR-100 show that SALT achieves higher accuracy than conventional retraining and fine-tuning while significantly reducing training cost. On CIFAR-10, SALT improves personalized accuracy from 88.1% to 93.8% while reducing training latency by more than 60%. SALT also maintains over 90% accuracy under 75% packet loss and preserves high accuracy (about 88% at sigma = 1.0) under noise injection. These results demonstrate that SALT provides an efficient and practical adaptation framework for real-world Split Computing systems.
Figures & tables
Fig. 1: Overview of practical challenges in Split Computing. Client devices execute a fixed head network and transmit intermediate representations to a server-side tail network under heterogeneous conditions. User-specific data distributions and varying communication quality can directly degrade inference performance, while transmitted representations may also be exposed to privacy threats such as inversion attacks in real-world edge–cloud systems.
Work
Primary Focus
Adaptation / Targeted Issue
Model Access Required
Neurosurgeon [ neurosurgeon ]
Split point scheduling
Latency / energy efficiency
Full
JointDNN [ jointdnn ]
Split training/inference framework
Latency / energy efficiency
Full
Split Computing & Early Exiting Survey [ sc_ee_survey ]
TABLE I: Summary of related work discussed in this work and their assumptions in split computing.
Fig. 2: Architecture of SALT, where a trainable adapter applies a correction to the latent feature produced by a frozen head network. The adapted feature is perturbed by noise and forwarded to a frozen tail network on the server. Only the adapter parameters are updated during training.
Fig. 3: Definition of split points in ResNet-18. The head network consists of the layers before the selected split point, while the tail network includes the remaining layers. We consider four split locations: BeforeBlock1, AfterBlock1, AfterBlock2, and AfterBlock3.
Configuration Item
Details
Optimizer
Adam
Loss Function
Cross Entropy Loss (on tail network outputs)
Learning Rate
1e−3
Batch Size
128
Training Epochs
100 (or until validation loss plateaus)
Activation Function
ReLU
TABLE II: Training configuration and environment used for all experiments.
Fig. 4: Illustration of the adaptation setting. While the model is pre-trained on a complete dataset, only a subset of these classes is used to simulate user-specific data.
Method
Adaptation Target
Parameter Access
Training Cost
Closed-Model Applicable
No Adaptation
None
–
None
✓
Head Retraining
Head Network
Full
High
×
Head Fine-Tuning
Head Network
Partial
Medium
×
SALT (Insertion)
Latent Features
Local Only
Medium–Low
✓
SALT (Adapter)
Latent Features
Local Only
Low
✓
TABLE III: Comparison of adaptation strategies in Split Computing.
Method
CIFAR-10
CIFAR-100
Accuracy
Epochs
Comp./Comm. Latency [s]
Inf. Latency [s]
Accuracy
Epochs
Comp./Comm. Latency [s]
Inf. Latency [s]
Original
0.881
-
- / -
5.3
0.669
-
- / -
1.0
Retrain
0.924
54.0
857.7 / 1110.2
5.4
0.824
49.6
122.1 / 208.3
1.0
Tune
0.920
26.0
372.3 / 534.6
5.4
0.889
39.6
101.3 / 166.3
1.0
SALT (Insertion Adapter)
0.934
25.3
287.6 / 520.2
5.3
0.902
27.2
60.7 / 114.2
1.0
SALT (Residual Adapter)
0.938
20.5
260.6 / 421.5
5.3
0.902
21.2
54.0 / 89.0
1.0
TABLE IV: Performance Comparison of Adaptation Methods on CIFAR-10 and CIFAR-100
Fig. 5: Accuracy vs. Total Training Latency for each method on CIFAR-10. SALT maintains high accuracy with low training cost. Error bars indicate 95% confidence intervals over 10 trials.
Fig. 6: Accuracy vs. Total Training Latency for each method on CIFAR-100. SALT maintains high accuracy with low training cost.
Fig. 7: Accuracy comparison across different split points on CIFAR-10. SALT consistently outperforms baseline methods regardless of split depth, achieving stable personalization without accessing or fine-tuning the head network.
Fig. 8: Accuracy under varying packet loss rates on CIFAR-10. While the original split model collapses under severe packet loss, SALT maintains stable and high accuracy across all loss rates without accessing the head or tail network parameters.
Fig. 9: Intermediate representations before packet loss under different loss rates. The figure shows channel-wise latent features (CH16) before feature dropping, for models trained with different packet loss rates. Higher loss rates lead to more spatially distributed representations, indicating redundancy-aware feature learning that improves robustness to packet loss.
Fig. 10: Visual comparison of inversion attack results under noise injection with and without SALT. The attacker is trained on noise-free representations and evaluated on noisy ones; reconstruction quality degrades similarly in both cases, indicating that SALT does not weaken noise-based privacy protection.
Fig. 11: Effect of noise strength on inversion attack performance (SSIM). SSIM decreases rapidly with increasing noise strength, and the nearly identical trends confirm that SALT does not degrade noise-based privacy defenses.
Fig. 12: Effect of noise strength on main-task classification accuracy. While noise injection severely degrades accuracy without adaptation, SALT maintains high accuracy across a wide range of noise strengths, demonstrating robustness to privacy-oriented perturbations.
Split Computing (SC) is a distributed inference method designed for balancing computational load and reducing latency by partitioning a neural network into head and tail networks and deploying them across both a mobile device and a server, respectively. Due to the limited bandwidth and computational power of mobile devices, the head network is often replaced with a neural network that requires lower computational and communication resources. However, it is challenging to prepare a sufficient quantity and quality of labeled data on the mobile device side for training the replacement head network. Furthermore, the data prepared by mobile users often contains sensitive privacy information, making data aggregation difficult. In this paper, we propose a data-efficient head network training method for SC. This method utilizes federated learning and knowledge distillation to train the head network using a small amount of non-IID and unlabeled data from mobile devices, ensuring that the training datasets are not exposed. Moreover, by introducing personalization, we enhance the accuracy to levels comparable to or exceeding those of existing methods. The experimental evaluations demonstrate that, despite data constraints imposed only on our method, it achieves accuracy comparable to the baselines. Furthermore, it significantly reduces the communication requirements by 95% on mobile devices, with only a 3-5 percentage point drop in accuracy from the original model.
Yutaro Horikawa, Takayuki Nishio
School of Engineering, Institute of Science Tokyo, Tokyo 152-8550, Japan
Adapting large AI models (LAMs) to personalized edge data is challenging because wireless devices have limited memory, computation, and uplink capacity. Federated fine-tuning preserves data privacy but still requires each device to host the full model, while split learning reduces device memory at the cost of heavy activation transmission. This paper proposes TSFLora, a token-compressed split fine-tuning framework for communication-efficient LAM adaptation at the edge. TSFLora combines attention-guided token selection, token merging, low-bit activation quantization, and LoRA-based adaptation within a split federated training pipeline. The key idea is to compress the intermediate token sequence before transmission so that the system reduces both uplink traffic and server-side processing without changing the frozen backbone. Experiments on ViT models over CIFAR-10, CIFAR-100, and TinyImageNet show that TSFLora achieves up to \textbf{6.8×} communication reduction and 41% memory saving while maintaining competitive accuracy.
Xianke Qiang, Zheng Chang, Li Wang +1
School of Computer Science and Engineering, UESTC, 611731 Chengdu, China · School of Computer Science, Beijing University of Posts and Telecommunications, Beijing, 100876, China. · Center for Intelligent Networking and Communications, UESTC, 611731 Chengdu, China.
Split learning (SL) enables collaborative training by partitioning a neural network across clients and a central server. However, the cut-layer interface introduces a key challenge: high-dimensional activations incur substantial communication overhead while exposing representations vulnerable to reconstruction attacks. Existing approaches typically address efficiency or privacy in isolation. We propose LightSplit, which limits information exposure and reduces communication overhead via a lightweight fixed orthogonal random projection at the cut layer. Based on Shannon's information theory, this projection acts as an information bottleneck that restricts instance-specific information and suppresses exploitable per-sample signals. With low-dimensional projections sent instead of raw activations, the server operates on lifted representations without requiring changes to existing SL architectures. Adding no trainable client components, LightSplit remains lightweight and suitable for edge devices while preserving end-to-end differentiability via exact gradient propagation. As the projection is non-invertible, part of the original representation is irreversibly discarded at the client, so LightSplit reduces the information available for reconstruction. We evaluate LightSplit on seven benchmarks spanning image classification, tabular data, and language-model fine-tuning, under IID and non-IID client data. At up to 32x compression it retains at least 94% of baseline accuracy on CIFAR-10, FashionMNIST, GTSRB, and UCI HAR, 86% on CIFAR-100, and 91% on average under non-IID client data; at 8x it retains 92% of ImageNet-1k top-1 and raises GPT-2 perplexity by at most 1.15x. Against state-of-the-art reconstruction attacks LightSplit resists inversion; the strongest, co-trained attack loses up to 87.3% of its reconstruction SSIM, and subject re-identification on UCI HAR drops by more than half.
Mert Cihangiroglu, Alessandro Pegoraro, Phillip Rieger +2
University of Pavia · Technical University of Darmstadt