Yue Niu

Sara Babakniya, Souvik Kundu, Saurav Prakash et al.

Edge devices can benefit remarkably from federated learning due to their distributed nature; however, their limited resource and computing power poses limitations in deployment. A possible solution to this problem is to utilize off-the-shelf sparse learning algorithms at the clients to meet their resource budget. However, such naive deployment in the clients causes significant accuracy degradation, especially for highly resource-constrained clients. In particular, our investigations reveal that the lack of consensus in the sparsity masks among the clients may potentially slow down the convergence of the global model and cause a substantial accuracy drop. With these observations, we present \textit{federated lottery aware sparsity hunting} (FLASH), a unified sparse learning framework for training a sparse sub-model that maintains the performance under ultra-low parameter density while yielding proportional communication benefits. Moreover, given that different clients may have different resource budgets, we present \textit{hetero-FLASH} where clients can take different density budgets based on their device resource limitations instead of supporting only one target parameter density. Experimental analysis on diverse models and datasets shows the superiority of FLASH in closing the gap with an unpruned baseline while yielding up to $\mathord{\sim}10.1\%$ improved accuracy with $\mathord{\sim}10.26\times$ fewer communication, compared to existing alternatives, at similar hyperparameter settings. Code is available at \url{https://github.com/SaraBabakN/flash_fl}.

Yue Niu, Saurav Prakash, Souvik Kundu et al.

Federated Learning (FL) is emerging as a popular, promising decentralized learning framework that enables collaborative training among clients, with no need to share private data between them or to a centralized server. However, considering many edge clients do not have sufficient computing, memory, or communication capabilities, federated learning of large models still faces significant bottlenecks. To keep such weak but crucial clients in the loop, prior works either consider a heterogeneous-client setting where clients train models with different sizes; or offload training to the server. However, the heterogeneous-client setting requires some clients to train full model, which is not aligned with the resource-constrained setting; while the latter ones break privacy promises in FL when sharing intermediate representations or labels with the server. To overcome these limitations, in this work, we formulate a realistic, but much less explored, cross-device FL setting in which no client can train a full large model nor is willing to share any intermediate information with the remote server. Under such a formulation, we develop a principal sub-model (PriSM) training methodology to collaboratively train a full large model, while assigning each client a small sub-model that is a probabilistic low-rank approximation to the full server model. When creating sub-models, PriSM first performs a principal kernel analysis in the orthogonal kernel space to obtain importance of each kernel. Then, PriSM adopts a novel importance-aware sampling process to select a subset of kernels (i.e., a kernel with high importance is assigned with a higher sampling probability). This sampling process ensures each sub-model is still a low-rank approximation to the full model, while all sub-models together achieve nearly full coverage on the principal kernels.

Yue Niu, Zalan Fabian, Sunwoo Lee et al.

Quasi-Newton methods still face significant challenges in training large-scale neural networks due to additional compute costs in the Hessian related computations and instability issues in stochastic training. A well-known method, L-BFGS that efficiently approximates the Hessian using history parameter and gradient changes, suffers convergence instability in stochastic training. So far, attempts that adapt L-BFGS to large-scale stochastic training incur considerable extra overhead, which offsets its convergence benefits in wall-clock time. In this paper, we propose mL-BFGS, a lightweight momentum-based L-BFGS algorithm that paves the way for quasi-Newton (QN) methods in large-scale distributed deep neural network (DNN) optimization. mL-BFGS introduces a nearly cost-free momentum scheme into L-BFGS update and greatly reduces stochastic noise in the Hessian, therefore stabilizing convergence during stochastic optimization. For model training at a large scale, mL-BFGS approximates a block-wise Hessian, thus enabling distributing compute and memory costs across all computing nodes. We provide a supporting convergence analysis for mL-BFGS in stochastic settings. To investigate mL-BFGS potential in large-scale DNN training, we train benchmark neural models using mL-BFGS and compare performance with baselines (SGD, Adam, and other quasi-Newton methods). Results show that mL-BFGS achieves both noticeable iteration-wise and wall-clock speedup.

Lei Gao, Amir Ziashahabi, Yue Niu et al.

Large Language Models (LLMs) are currently pre-trained and fine-tuned on large cloud servers. The next frontier is LLM personalization, where a foundation model can be fine-tuned with user/task-specific data. Given the sensitive nature of such private data, it is desirable to fine-tune these models on edge devices to improve user trust. However, fine-tuning on resource-constrained edge devices presents significant challenges due to substantial memory and computational demands, as well as limited infrastructure support. We observe that inference engines (e.g., ExecuTorch) can be repurposed for fine-tuning by leveraging zeroth-order (ZO) optimization, which uses multiple forward passes to approximate gradients. While promising, direct application of ZO methods on edge devices is inefficient due to the high computational cost of multiple forward passes required for accurate gradient estimation, and their deployment has been largely unexplored in practice. We introduce MobiZO, a resource-efficient fine-tuning framework for LLMs specifically designed for edge devices. MobiZO combines three key innovations: (1) a parallelized randomized gradient estimator that employs both outer-loop and inner-loop parallelism to eliminate sequential forward passes, (2) a specialized Multi-Perturbed LoRA (MP-LoRA) module that enables efficient realization of both inner and outer loop parallelism, and (3) a seamless integration with ExecuTorch for on-device training, requiring no modifications to the runtime. Experiments demonstrate that MobiZO achieves substantial runtime speedups and memory savings while improving fine-tuning accuracy, paving the way for practical deployment of LLMs in real-time, on-device applications.

Lei Gao, Yue Niu, Tingting Tang et al.

Language models (LMs) have greatly propelled the research on natural language processing. However, LMs also raise concerns regarding the generation of biased or toxic content and the potential disclosure of private information from the training dataset. In this work, we present a new efficient approach, Ethos, that rectifies LMs to mitigate toxicity and bias in outputs and avoid privacy leakage. Ethos is built on task arithmetic. However, unlike current task arithmetic algorithms, Ethos distinguishes general beneficial and undesired knowledge when reconstructing task vectors. Specifically, Ethos first obtains a set of principal components from the pre-trained models using singular value decomposition. Then, by projecting the task vector onto principal components, Ethos identifies the principal components that encode general or undesired knowledge. Ethos performs negating using the task vector with undesired knowledge only, thereby minimizing collateral damage on general model utility. We demonstrate the efficacy of our approach on three different tasks: debiasing, detoxification, and memorization unlearning. Evaluations show Ethos is more effective in removing undesired knowledge and maintaining the overall model performance compared to current task arithmetic methods.

Tingting Tang, Yue Niu, Salman Avestimehr et al.

Graph neural networks (GNNs) play a key role in learning representations from graph-structured data and are demonstrated to be useful in many applications. However, the GNN training pipeline has been shown to be vulnerable to node feature leakage and edge extraction attacks. This paper investigates a scenario where an attacker aims to recover private edge information from a trained GNN model. Previous studies have employed differential privacy (DP) to add noise directly to the adjacency matrix or a compact graph representation. The added perturbations cause the graph structure to be substantially morphed, reducing the model utility. We propose a new privacy-preserving GNN training algorithm, Eclipse, that maintains good model utility while providing strong privacy protection on edges. Eclipse is based on two key observations. First, adjacency matrices in graph structures exhibit low-rank behavior. Thus, Eclipse trains GNNs with a low-rank format of the graph via singular values decomposition (SVD), rather than the original graph. Using the low-rank format, Eclipse preserves the primary graph topology and removes the remaining residual edges. Eclipse adds noise to the low-rank singular values instead of the entire graph, thereby preserving the graph privacy while still maintaining enough of the graph structure to maintain model utility. We theoretically show Eclipse provide formal DP guarantee on edges. Experiments on benchmark graph datasets show that Eclipse achieves significantly better privacy-utility tradeoff compared to existing privacy-preserving GNN training methods. In particular, under strong privacy constraints ($ε$ < 4), Eclipse shows significant gains in the model utility by up to 46%. We further demonstrate that Eclipse also has better resilience against common edge attacks (e.g., LPA), lowering the attack AUC by up to 5% compared to other state-of-the-art baselines.

Yue Niu, Ramy E. Ali, Saurav Prakash et al.

Data privacy is of great concern in cloud machine-learning service platforms, when sensitive data are exposed to service providers. While private computing environments (e.g., secure enclaves), and cryptographic approaches (e.g., homomorphic encryption) provide strong privacy protection, their computing performance still falls short compared to cloud GPUs. To achieve privacy protection with high computing performance, we propose Delta, a new private training and inference framework, with comparable model performance as non-private centralized training. Delta features two asymmetric data flows: the main information-sensitive flow and the residual flow. The main part flows into a small model while the residuals are offloaded to a large model. Specifically, Delta embeds the information-sensitive representations into a low-dimensional space while pushing the information-insensitive part into high-dimension residuals. To ensure privacy protection, the low-dimensional information-sensitive part is secured and fed to a small model in a private environment. On the other hand, the residual part is sent to fast cloud GPUs, and processed by a large model. To further enhance privacy and reduce the communication cost, Delta applies a random binary quantization technique along with a DP-based technique to the residuals before sharing them with the public platform. We theoretically show that Delta guarantees differential privacy in the public environment and greatly reduces the complexity in the private environment. We conduct empirical analyses on CIFAR-10, CIFAR-100 and ImageNet datasets and ResNet-18 and ResNet-34, showing that Delta achieves strong privacy protection, fast training, and inference without significantly compromising the model utility.

Yue Niu, Saurav Prakash, Salman Avestimehr

We propose a new attention mechanism with linear complexity, ATP, that fixates \textbf{A}ttention on \textbf{T}op \textbf{P}rincipal keys, rather than on each individual token. Particularly, ATP is driven by an important observation that input sequences are typically low-rank, i.e., input sequences can be represented by a few principal bases. Therefore, instead of directly iterating over all the input tokens, ATP transforms inputs into an orthogonal space and computes attention only on the top principal bases (keys). Owing to the observed low-rank structure in input sequences, ATP is able to capture semantic relationships in input sequences with a few principal keys. Furthermore, the attention complexity is reduced from \emph{quadratic} to \emph{linear} without incurring a noticeable performance drop. ATP further reduces complexity for other linear layers with low-rank inputs, leading to more speedup compared to prior works that solely target the attention module. Our evaluations on various models (e.g., BERT and Llama) demonstrate that ATP achieves comparable accuracy with much lower computation and memory complexity than the standard attention mechanism. In particular, ATP barely loses accuracy with only $1/2$ principal keys, and only incurs around $2\%$ accuracy drops with $1/4$ principal keys.

Sunwoo Lee, Tuo Zhang, Saurav Prakash et al.

In Federated Learning (FL), clients may have weak devices that cannot train the full model or even hold it in their memory space. To implement large-scale FL applications, thus, it is crucial to develop a distributed learning method that enables the participation of such weak clients. We propose EmbracingFL, a general FL framework that allows all available clients to join the distributed training regardless of their system resource capacity. The framework is built upon a novel form of partial model training method in which each client trains as many consecutive output-side layers as its system resources allow. Our study demonstrates that EmbracingFL encourages each layer to have similar data representations across clients, improving FL efficiency. The proposed partial model training method guarantees convergence to a neighbor of stationary points for non-convex and smooth problems. We evaluate the efficacy of EmbracingFL under a variety of settings with a mixed number of strong, moderate (~40% memory), and weak (~15% memory) clients, datasets (CIFAR-10, FEMNIST, and IMDB), and models (ResNet20, CNN, and LSTM). Our empirical study shows that EmbracingFL consistently achieves high accuracy as like all clients are strong, outperforming the state-of-the-art width reduction methods (i.e. HeteroFL and FjORD).

Yue Niu, Ramy E. Ali, Salman Avestimehr

Leveraging parallel hardware (e.g. GPUs) for deep neural network (DNN) training brings high computing performance. However, it raises data privacy concerns as GPUs lack a trusted environment to protect the data. Trusted execution environments (TEEs) have emerged as a promising solution to achieve privacy-preserving learning. Unfortunately, TEEs' limited computing power renders them not comparable to GPUs in performance. To improve the trade-off among privacy, computing performance, and model accuracy, we propose an \emph{asymmetric} model decomposition framework, \AsymML{}, to (1) accelerate training using parallel hardware; and (2) achieve a strong privacy guarantee using TEEs and differential privacy (DP) with much less accuracy compromised compared to DP-only methods. By exploiting the low-rank characteristics in training data and intermediate features, \AsymML{} asymmetrically decomposes inputs and intermediate activations into low-rank and residual parts. With the decomposed data, the target DNN model is accordingly split into a \emph{trusted} and an \emph{untrusted} part. The trusted part performs computations on low-rank data, with low compute and memory costs. The untrusted part is fed with residuals perturbed by very small noise. Privacy, computing performance, and model accuracy are well managed by respectively delegating the trusted and the untrusted part to TEEs and GPUs. We provide a formal DP guarantee that demonstrates that, for the same privacy guarantee, combining asymmetric data decomposition and DP requires much smaller noise compared to solely using DP without decomposition. This improves the privacy-utility trade-off significantly compared to using only DP methods without decomposition. Furthermore, we present a rank bound analysis showing that the low-rank structure is preserved after each layer across the entire model.

Yue Niu, Hanqing Zeng, Ajitesh Srivastava et al.

To accelerate inference of Convolutional Neural Networks (CNNs), various techniques have been proposed to reduce computation redundancy. Converting convolutional layers into frequency domain significantly reduces the computation complexity of the sliding window operations in space domain. On the other hand, weight pruning techniques address the redundancy in model parameters by converting dense convolutional kernels into sparse ones. To obtain high-throughput FPGA implementation, we propose SPEC2 -- the first work to prune and accelerate spectral CNNs. First, we propose a systematic pruning algorithm based on Alternative Direction Method of Multipliers (ADMM). The offline pruning iteratively sets the majority of spectral weights to zero, without using any handcrafted heuristics. Then, we design an optimized pipeline architecture on FPGA that has efficient random access into the sparse kernels and exploits various dimensions of parallelism in convolutional layers. Overall, SPEC2 achieves high inference throughput with extremely low computation complexity and negligible accuracy degradation. We demonstrate SPEC2 by pruning and implementing LeNet and VGG16 on the Xilinx Virtex platform. After pruning 75% of the spectral weights, SPEC2 achieves 0% accuracy loss for LeNet, and <1% accuracy loss for VGG16. The resulting accelerators achieve up to 24x higher throughput, compared with the state-of-the-art FPGA implementations for VGG16.