Reassessing Layer Pruning in LLMs: New Insights and Methods

Although large language models (LLMs) have achieved remarkable success across various domains, their considerable scale necessitates substantial computational resources, posing significant challenges for deployment in resource-constrained environments. Layer pruning, as a simple yet effective compression method, removes layers of a model directly, reducing computational overhead. However, what are the best practices for layer pruning in LLMs? Are sophisticated layer selection metrics truly effective? Does the LoRA (Low-Rank Approximation) family, widely regarded as a leading method for pruned model fine-tuning, truly meet expectations when applied to post-pruning fine-tuning? To answer these questions, we dedicate thousands of GPU hours to benchmarking layer pruning in LLMs and gaining insights across multiple dimensions. Our results demonstrate that a simple approach, i.e., pruning the final 25\% of layers followed by fine-tuning the \texttt{lm\_head} and the remaining last three layer, yields remarkably strong performance. Following this guide, we prune Llama-3.1-8B-It and obtain a model that outperforms many popular LLMs of similar size, such as ChatGLM2-6B, Vicuna-7B-v1.5, Qwen1.5-7B and Baichuan2-7B. We release the optimal model weights on Huggingface, and the code is available on GitHub.

RedTest: Towards Measuring Redundancy in Deep Neural Networks Effectively

Deep learning has revolutionized computing in many real-world applications, arguably due to its remarkable performance and extreme convenience as an end-to-end solution. However, deep learning models can be costly to train and to use, especially for those large-scale models, making it necessary to optimize the original overly complicated models into smaller ones in scenarios with limited resources such as mobile applications or simply for resource saving. The key question in such model optimization is, how can we effectively identify and measure the redundancy in a deep learning model structure. While several common metrics exist in the popular model optimization techniques to measure the performance of models after optimization, they are not able to quantitatively inform the degree of remaining redundancy. To address the problem, we present a novel testing approach, i.e., RedTest, which proposes a novel testing metric called Model Structural Redundancy Score (MSRS) to quantitatively measure the degree of redundancy in a deep learning model structure. We first show that MSRS is effective in both revealing and assessing the redundancy issues in many state-of-the-art models, which urgently calls for model optimization. Then, we utilize MSRS to assist deep learning model developers in two practical application scenarios: 1) in Neural Architecture Search, we design a novel redundancy-aware algorithm to guide the search for the optimal model structure and demonstrate its effectiveness by comparing it to existing standard NAS practice; 2) in the pruning of large-scale pre-trained models, we prune the redundant layers of pre-trained models with the guidance of layer similarity to derive less redundant ones of much smaller size. Extensive experimental results demonstrate that removing such redundancy has a negligible effect on the model utility.

Lateral Movement Detection via Time-aware Subgraph Classification on Authentication Logs

Lateral movement is a crucial component of advanced persistent threat (APT) attacks in networks. Attackers exploit security vulnerabilities in internal networks or IoT devices, expanding their control after initial infiltration to steal sensitive data or carry out other malicious activities, posing a serious threat to system security. Existing research suggests that attackers generally employ seemingly unrelated operations to mask their malicious intentions, thereby evading existing lateral movement detection methods and hiding their intrusion traces. In this regard, we analyze host authentication log data from a graph perspective and propose a multi-scale lateral movement detection framework called LMDetect. The main workflow of this framework proceeds as follows: 1) Construct a heterogeneous multigraph from host authentication log data to strengthen the correlations among internal system entities; 2) Design a time-aware subgraph generator to extract subgraphs centered on authentication events from the heterogeneous authentication multigraph; 3) Design a multi-scale attention encoder that leverages both local and global attention to capture hidden anomalous behavior patterns in the authentication subgraphs, thereby achieving lateral movement detection. Extensive experiments on two real-world authentication log datasets demonstrate the effectiveness and superiority of our framework in detecting lateral movement behaviors.

Rethinking Graph Transformer Architecture Design for Node Classification

Graph Transformer (GT), as a special type of Graph Neural Networks (GNNs), utilizes multi-head attention to facilitate high-order message passing. However, this also imposes several limitations in node classification applications: 1) nodes are susceptible to global noise; 2) self-attention computation cannot scale well to large graphs. In this work, we conduct extensive observational experiments to explore the adaptability of the GT architecture in node classification tasks and draw several conclusions: the current multi-head self-attention module in GT can be completely replaceable, while the feed-forward neural network module proves to be valuable. Based on this, we decouple the propagation (P) and transformation (T) of GNNs and explore a powerful GT architecture, named GNNFormer, which is based on the P/T combination message passing and adapted for node classification in both homophilous and heterophilous scenarios. Extensive experiments on 12 benchmark datasets demonstrate that our proposed GT architecture can effectively adapt to node classification tasks without being affected by global noise and computational efficiency limitations.

MDM: Advancing Multi-Domain Distribution Matching for Automatic Modulation Recognition Dataset Synthesis

Recently, deep learning technology has been successfully introduced into Automatic Modulation Recognition (AMR) tasks. However, the success of deep learning is all attributed to the training on large-scale datasets. Such a large amount of data brings huge pressure on storage, transmission and model training. In order to solve the problem of large amount of data, some researchers put forward the method of data distillation, which aims to compress large training data into smaller synthetic datasets to maintain its performance. While numerous data distillation techniques have been developed within the realm of image processing, the unique characteristics of signals set them apart. Signals exhibit distinct features across various domains, necessitating specialized approaches for their analysis and processing. To this end, a novel dataset distillation method--Multi-domain Distribution Matching (MDM) is proposed. MDM employs the Discrete Fourier Transform (DFT) to translate timedomain signals into the frequency domain, and then uses a model to compute distribution matching losses between the synthetic and real datasets, considering both the time and frequency domains. Ultimately, these two losses are integrated to update the synthetic dataset. We conduct extensive experiments on three AMR datasets. Experimental results show that, compared with baseline methods, our method achieves better performance under the same compression ratio. Furthermore, we conduct crossarchitecture generalization experiments on several models, and the experimental results show that our synthetic datasets can generalize well on other unseen models.

Enhancing Ethereum Fraud Detection via Generative and Contrastive Self-supervision

The rampant fraudulent activities on Ethereum hinder the healthy development of the blockchain ecosystem, necessitating the reinforcement of regulations. However, multiple imbalances involving account interaction frequencies and interaction types in the Ethereum transaction environment pose significant challenges to data mining-based fraud detection research. To address this, we first propose the concept of meta-interactions to refine interaction behaviors in Ethereum, and based on this, we present a dual self-supervision enhanced Ethereum fraud detection framework, named Meta-IFD. This framework initially introduces a generative self-supervision mechanism to augment the interaction features of accounts, followed by a contrastive self-supervision mechanism to differentiate various behavior patterns, and ultimately characterizes the behavioral representations of accounts and mines potential fraud risks through multi-view interaction feature learning. Extensive experiments on real Ethereum datasets demonstrate the effectiveness and superiority of our framework in detecting common Ethereum fraud behaviors such as Ponzi schemes and phishing scams. Additionally, the generative module can effectively alleviate the interaction distribution imbalance in Ethereum data, while the contrastive module significantly enhances the framework's ability to distinguish different behavior patterns. The source code will be released on GitHub soon.

A Generic Layer Pruning Method for Signal Modulation Recognition Deep Learning Models

With the successful application of deep learning in communications systems, deep neural networks are becoming the preferred method for signal classification. Although these models yield impressive results, they often come with high computational complexity and large model sizes, which hinders their practical deployment in communication systems. To address this challenge, we propose a novel layer pruning method. Specifically, we decompose the model into several consecutive blocks, each containing consecutive layers with similar semantics. Then, we identify layers that need to be preserved within each block based on their contribution. Finally, we reassemble the pruned blocks and fine-tune the compact model. Extensive experiments on five datasets demonstrate the efficiency and effectiveness of our method over a variety of state-of-the-art baselines, including layer pruning and channel pruning methods.

Deep Image Semantic Communication Model for Artificial Intelligent Internet of Things

With the rapid development of Artificial Intelligent Internet of Things (AIoT), the image data from AIoT devices has been witnessing the explosive increasing. In this paper, a novel deep image semantic communication model is proposed for the efficient image communication in AIoT. Particularly, at the transmitter side, a high-precision image semantic segmentation algorithm is proposed to extract the semantic information of the image to achieve significant compression of the image data. At the receiver side, a semantic image restoration algorithm based on Generative Adversarial Network (GAN) is proposed to convert the semantic image to a real scene image with detailed information. Simulation results demonstrate that the proposed image semantic communication model can improve the image compression ratio and recovery accuracy by 71.93% and 25.07% on average in comparison with WebP and CycleGAN, respectively. More importantly, our demo experiment shows that the proposed model reduces the total delay by 95.26% in the image communication, when comparing with the original image transmission.

Augmenting Radio Signals with Wavelet Transform for Deep Learning-Based Modulation Recognition

The use of deep learning for radio modulation recognition has become prevalent in recent years. This approach automatically extracts high-dimensional features from large datasets, facilitating the accurate classification of modulation schemes. However, in real-world scenarios, it may not be feasible to gather sufficient training data in advance. Data augmentation is a method used to increase the diversity and quantity of training dataset and to reduce data sparsity and imbalance. In this paper, we propose data augmentation methods that involve replacing detail coefficients decomposed by discrete wavelet transform for reconstructing to generate new samples and expand the training set. Different generation methods are used to generate replacement sequences. Simulation results indicate that our proposed methods significantly outperform the other augmentation methods.

Can pre-trained models assist in dataset distillation?

Dataset Distillation (DD) is a prominent technique that encapsulates knowledge from a large-scale original dataset into a small synthetic dataset for efficient training. Meanwhile, Pre-trained Models (PTMs) function as knowledge repositories, containing extensive information from the original dataset. This naturally raises a question: Can PTMs effectively transfer knowledge to synthetic datasets, guiding DD accurately? To this end, we conduct preliminary experiments, confirming the contribution of PTMs to DD. Afterwards, we systematically study different options in PTMs, including initialization parameters, model architecture, training epoch and domain knowledge, revealing that: 1) Increasing model diversity enhances the performance of synthetic datasets; 2) Sub-optimal models can also assist in DD and outperform well-trained ones in certain cases; 3) Domain-specific PTMs are not mandatory for DD, but a reasonable domain match is crucial. Finally, by selecting optimal options, we significantly improve the cross-architecture generalization over baseline DD methods. We hope our work will facilitate researchers to develop better DD techniques. Our code is available at https://github.com/yaolu-zjut/DDInterpreter.