In the ever-evolving landscape of Deep Neural Networks (DNN) hardware acceleration, unlocking the true potential of systolic array accelerators has long been hindered by the daunting challenges of expertise and time investment. Large Language Models (LLMs) offer a promising solution for automating code generation, which is key to unlocking unprecedented efficiency and performance in various domains, including hardware descriptive code. The generative power of LLMs can enable the effective utilization of preexisting designs and dedicated hardware generators. However, the successful application of LLMs to hardware accelerator design is contingent upon the availability of specialized datasets tailored for this purpose. To bridge this gap, we introduce the Systolic Array-based Accelerator DataSet (SA-DS). SA-DS comprises a diverse collection of spatial array designs following the standardized Berkeley’s Gemmini accelerator generator template, enabling design reuse, adaptation, and customization. SA-DS is intended to spark LLM-centered research on DNN hardware accelerator architecture. We envision that SA-DS provides a framework that will shape the course of DNN hardware acceleration research for generations to come. SA-DS is open-sourced under the permissive MIT license at https://github.com/ACADLab/SA-DS.

The capability of generating high-quality source code using large language models (LLMs) reduces software development time and costs. However, they often introduce security vulnerabilities due to training on insecure open-source data. This highlights the need for ensuring secure and functional code generation. This paper introduces PromSec, an algorithm for prom optimization for secure and functioning code generation using LLMs. In PromSec, we combine 1) code vulnerability clearing using a generative adversarial graph neural network, dubbed as gGAN, to fix and reduce security vulnerabilities in generated codes and 2) code generation using an LLM into an interactive loop, such that the outcome of the gGAN drives the LLM with enhanced prompts to generate secure codes while preserving their functionality. Introducing a new contrastive learning approach in gGAN, we formulate code-clearing and generation as a dual-objective optimization problem, enabling PromSec to notably reduce the number of LLM inferences. PromSec offers a cost-effective and practical solution for generating secure, functional code. Extensive experiments conducted on Python and Java code datasets confirm that PromSec effectively enhances code security while upholding its intended functionality. Our experiments show that while a state-of-the-art approach fails to address all code vulnerabilities, PromSec effectively resolves them. Moreover, PromSec achieves more than an order-of-magnitude reduction in operation time, number of LLM queries, and security analysis costs. Furthermore, prompts optimized with PromSec for a certain LLM are transferable to other LLMs across programming languages and generalizable to unseen vulnerabilities in training. This study is a step in enhancing the trustworthiness of LLMs for secure and functional code generation, supporting their integration into real-world software development.

This paper introduces SGCode, a flexible prompt-optimizing system to generate secure code with large language models (LLMs). SGCode integrates recent prompt-optimization approaches with LLMs in a unified system accessible through front-end and back-end APIs, enabling users to 1) generate secure code, which is free of vulnerabilities, 2) review and share security analysis, and 3) easily switch from one prompt optimization approach to another, while providing insights on model and system performance. We populated SGCode on an AWS server with PromSec, an approach that optimizes prompts by combining an LLM and security tools with a lightweight generative adversarial graph neural network to detect and fix security vulnerabilities in the generated code. Extensive experiments show that SGCode is practical as a public tool to gain insights into the trade-offs between model utility, secure code generation, and system cost. SGCode has only a marginal cost compared with prompting LLMs. SGCode is available at: http://3.131.141.63:8501/.

Malicious domain detection (MDD) is an open security challenge that aims to detect if an Internet domain name is associated with cyber attacks. Many techniques have been applied to tackle this problem, among which graph neural networks (GNNs) are deemed one of the most effective approaches. GNN-based MDD employs domain name system (DNS) logs to represent Internet domains as nodes in a graph, dubbed domain maliciousness graph (DMG) and trains a GNN model to infer the maliciousness of Internet domains by leveraging the maliciousness of already identified ones. As this method heavily relies on the "publicly" accessible DNS logs to build DMGs, it creates a vulnerability for adversaries to manipulate the features and edges of their domain nodes within these graphs. The current body of literature primarily focuses on threat models that involve manipulating individual adversary (attacker) nodes. Nonetheless, adversaries usually create numerous domains to accomplish their attack objectives, aiming to reduce costs and evade detection. Hence, they aim to remain undetected across as many domains as possible. In this work, we call the attack that manipulates several nodes in the DMG concurrently a multi-instance evasion attack. To the best of our knowledge, this type of attack has not been explored in the prior art. We present both theoretical and empirical evidence to show that the existing single-instance evasion techniques for GNN-based MDDs are inadequate to launch multi-instance evasion attacks. Therefore, we propose an inference-time, multi-instance adversarial attack, dubbed MintA, against GNN-based MDD. MintA optimizes node perturbations to enhance the evasiveness of a node and its neighborhood. MintA only requires black-box access to the target model to launch the attack successfully. In other words, MintA does not require any knowledge of the MDD model’s parameters, architecture, or information on non-adversary nodes. We formulate an optimization problem that satisfies the attack objectives of MintA and devise an approximate solution for it. We evaluate MintA on a state-of-the-art GNN-based MDD technique using real-world data, and our experiments demonstrate an attack success rate of over 80%. The findings of this study serve as a cautionary note for security experts, highlighting the vulnerability of GNN-based MDD to practical attacks that can impede the effectiveness and advantages of this approach.

In this paper, we propose IMA-GNN as an In-Memory Accelerator for centralized and decentralized Graph Neural Network inference, explore its potential in both settings and provide a guideline for the community targeting flexible and efficient edge computation. Leveraging IMA-GNN, we first model the computation and communication latencies of edge devices. We then present practical case studies on GNN-based taxi demand and supply prediction and also adopt four large graph datasets to quantitatively compare and analyze centralized and decentralized settings. Our cross-layer simulation results demonstrate that on average, IMA-GNN in the centralized setting can obtain ~790x communication speed-up compared to the decentralized GNN setting. However, the decentralized setting performs computation ~1400x faster while reducing the power consumption per device. This further underlines the need for a hybrid semi-decentralized GNN approach.

Identifying spectrum opportunities is a crucial element of efficient spectrum utilization for future wireless networks. Spectrum sensing offers a convenient means for revealing such opportunities. Studies showed that usage of the spectrum has a high correlation over multi-dimensions, including time and frequency. However, multi-dimensional spectrum sensing requires high-cost processes. Applying compressive sensing allows for subNyquist sampling. This reduces associated training, feedback, and computation overheads of a spectrum sensing method. However, the accuracy of the signal sparsity assumption and knowledge of the precise sparsity level are necessary for the applicability of compressive sensing. It is common practice to assume a level of known sparsity. On the other hand, in reality, this presumption is incorrect. This paper proposes a method for estimating the multidimensional sparsity for spectrum sensing. By extrapolating it from its counterpart with respect to a compact discrete Fourier basis, the proposed method calculates the sparsity level over a dictionary. A machine learning estimation method achieves this inference. Extensive simulations validate a high-quality sparsity estimation. To validate this observation, real-world measurements are used, where one of the biggest Turkish telecom operators has private uplink bands in the frequency range between 852-856 MHz.

The growing use of media has led to the development of several machine learning (ML) and natural language processing (NLP) tools to process the unprecedented amount of social media content to make actionable decisions. However, these ML and NLP algorithms have been widely shown to be vulnerable to adversarial attacks. These vulnerabilities allow adversaries to launch a diversified set of adversarial attacks on these algorithms in different applications of social media text processing. In this article, we provide a comprehensive review of the main approaches for adversarial attacks and defenses in the context of social media applications with a particular focus on key challenges and future research directions. In detail, we cover literature on six key applications: 1) rumors detection; 2) satires detection; 3) clickbaits and spams identification; 4) hate speech detection; 5) misinformation detection; and 6) sentiment analysis. We then highlight the concurrent and anticipated future research questions and provide recommendations and directions for future work.

In cognitive radio systems, predicting spectrum occupancies is a convenient alternative way to continuous spectrum sensing. It can provide information on spectrum usage and so empty spectrum bands can be used by secondary users. The usage of the spectrum bands is highly correlated over both time and frequency. Recently, machine learning algorithms are used to predict spectrum occupancy by exploiting such correlations. However, this approach primarily assumes a supervised learning setting. Despite its outstanding performance, this setting requires the availability of sufficiently large datasets (of labeled data) and is not adaptive to environment changes. In this paper, different from the existing literature, a deep reinforcement learning (RL) algorithm is used to alleviate those shortcomings. In this algorithm, we define the reward functions of the deep RL setting and its state and action spaces such that it is applicable to work dynamically, in an online fashion, in real world settings. Extensive experiments validate the capability of the proposed algorithm in predicting spectrum occupancies as examined over real world spectrum measurements. These are carried out in the 832-862 megahertz frequency bands, which are used by the leading Turkish telecom providers as private uplink bands. This is a significant step towards realizing a standalone spectrum occupancy prediction operation without any control from the operator and minimizing memory requirements while alleviating the need for the labeled dataset.

Obtaining accurate channel state estimates at reasonable training overheads remains a big challenge for the applicability of multiple-input multiple-output orthogonal frequency division multiplexing (MIMO-OFDM). Recently, the exploitation of channel sparsity has led to sub-Nyquist channel sampling thereby reducing the channel training overhead. Still, there is a growing belief in channel sparsity appearance in many dimensions; time, frequency, angle, and space. Accordingly, this paper proposes an algorithm for channel estimation where sparsity in multidimensions is simultaneously exploited. Also, the applicability of sparse coding relies on the validity of a signal sparsity assumption and knowing the exact sparsity level. However, this assumption is not valid in practice, especially when applying learned dictionaries as sparsifying transforms. The problem is more strongly pronounced with multidimensional sparsity. In this paper, we also propose an algorithm for estimating the composite sparsity lying in multiple domains defined by learned dictionaries. Simulations validate a substantial channel estimation quality attained by the proposed algorithm as compared to the existing algorithms. The simulations also validate a high quality of sparsity estimation leading to performances close to the impractical case of assuming known sparsity.

In sparse coding-based channel estimation, atom selection is based on jointly minimizing the sparsity and the error of the representation of the noisy measurement. However, this selection is not necessarily optimal in terms of minimizing the channel estimation error. This calls for better ways of atom selection. Accordingly, we propose an algorithm for improved atom selection in sparse coding for frequency division duplex (FDD) massive multiple-input-multiple-output (MIMO) downlink channel estimation. The proposed algorithm performs iterative atom selection based on two residuals. First is the received signal residual used to guide on a small selection pool of candidate atoms. Second is the residual of an initial channel estimate which is used to pick the best atom within the selection pool. Simulation results show the advantage of the proposed algorithm over standard sparse coding-based channel estimation. Moreover, the proposed algorithm eliminates the need for cell-specific trained dictionaries without sacrificing the performance. Furthermore, the proposed sparse coding can be applied in the process of dictionary learning to train for improved dictionaries achieving further performance enhancement.

The reconfigurable intelligent surface (RIS) technology has attracted interest due to its promising coverage and spectral efficiency features. However, some challenges need to be addressed to realize this technology in practice. One of the main challenges is the configuration of reflecting coefficients without the need for beam training overhead or massive channel estimation. Earlier works used estimated channel information with deep learning algorithms to design RIS reflection matrices. Although these works can reduce the beam training overhead, still they overlook existing correlations in the previously sampled channels. In this paper, different from existing works, we propose to exploit the correlation in the previously sampled channels to estimate RIS interaction more reliably. We use a deep multilayer perceptron for this purpose. Simulation results reveal performance improvements achieved by the proposed algorithm.

Cognitive radio (CR)-based internet of things systems can be considered as an efficient solution for futuristic smart technologies. However, CRs are naturally vulnerable to two major security threats; primary user emulation (PUE) and jamming attacks. Machine learning has been recently applied to the detection of these attacks. Still, the need for feature extraction required by machine learning techniques restrains the full exploitation of raw data. To alleviate this need, this paper proposes one-dimensional deep learning as a framework for identifying such attacks. Simulations show the ability of the proposed algorithm to detect these attacks with high performance.

The identification of spectrum opportunities is a pivotal requirement for efficient spectrum utilization in cognitive radio systems. Spectrum prediction offers a convenient means for revealing such opportunities based on the previously obtained occupancies. As spectrum occupancy states are correlated over time, spectrum prediction is often cast as a predictable time-series process using classical or deep learning-based models. However, this variety of methods exploits time-domain correlation and overlooks the existing correlation over frequency. In this paper, differently from previous works, we investigate a more realistic scenario by exploiting correlation over time and frequency through a 2D-long short-term memory (LSTM) model. Extensive experimental results show a performance improvement over conventional spectrum prediction methods in terms of accuracy and computational complexity. These observations are validated over the real-world spectrum measurements, assuming a frequency range between 832-862 MHz where most of the telecom operators in Turkey have private uplink bands.

Recent research considers the application of a lens antenna array in order to provide efficient beam selection in beamspace massive MIMO. Achieving the advantages of this beam selection paradigm requires efficient channel estimation in the beamspace. Along this line, beamspace sparsity is an efficient regularizer to this problem. In this paper, we propose using a dictionary trained over a set of example beam selection matrices, as a beam selection tool. In this context, a learned dictionary can more effectively guarantee the sparsity of the representation at the specified sparsity level, owing to the dictionary learning process. This means that it gives a better sparse representation, and, consequently, a better channel estimation quality. Simulations validate that using a trained dictionary improves the quality of channel estimation, as tested over two channel models with different operating scenarios.

Despite the well-known success of sub-Nyquist sampling in reducing the hardware and computational costs of spectrum sensing, it still has the shortcoming of requiring a pre-determined spectrum sparsity level. This paper proposes an algorithm for sub-Nyquist wide-band spectrum sensing addressing this shortcoming. The proposed algorithm divides the spectrum into narrow, contagious frequency subbands and learns a subband dictionary for each subband. A subband dictionary is well-suited for the representation of signals in its corresponding subband. A compressed version of the received signal is sparsely coded over each subband dictionary. We show that the convergence patterns over a specific dictionary can be used for identifying the occupancy of its underlying subband. Therefore, the convergence patterns obtained by the gradient operator are used as distinctive classifying features. Then, a machine learning-based classifier is trained over these features and used to make the decision about spectrum occupancy. As the interest is only to characterize sparse coding convergence patterns, we alleviate the need for a specific or an estimated sparsity level. Besides, using subband dictionaries at different frequencies omits the need for a frequency-splitting filterbank. The proposed algorithm achieves significant performance improvements in terms of the probability-of-detection and false-alarm-rate measures. This result is validated through simulations with various operating scenarios.

Sparse coding over a redundant dictionary has recently been used as a framework for downlink channel estimation in frequency division duplex massive multiple-input multiple-output antenna systems. This usage allows for efficiently reducing the inherently high training and feedback overheads. We present an algorithm for downlink channel estimation via selective sparse coding over multiple cluster dictionaries. A channel training set is divided into clusters based on the angle of the arrival/departure of the majority physical subpaths corresponding to each channel tap. Then, a compact dictionary is trained in each cluster. Channel estimation is done by first identifying the channel cluster and then using its dictionary for reconstruction. This selective sparse coding allows for adaptive regularization via sparse model selection, thereby offering additional regularization to the ill-posed channel estimation problem. We empirically validate the selectivity of the cluster dictionaries. Simulation results show the advantage of the proposed algorithm in achieving better estimation quality at lower computational cost, as compared the case of using standard sparse coding.

A new strategy for multiple structured dictionary learning is proposed. It is motivated by the fact that a signal and its residual after sparse approximation do not necessarily possess the same geometric structure. Based on the geometric structure of each residual component, the most appropriate dictionary is selected. A single-atom sparse representation vector of this residual is calculated and the chosen dictionary is updated. For a given training signal, the process of model (dictionary) selection and one-atom representation is repeated until the desired sparsity or approximation error is reached. Thus, the proposed strategy provides a mechanism whereby each signal can update the most relevant dictionaries based on the structure of its residuals. Simulations conducted over natural images show that, in comparison to standard single or multiple dictionary learning and sparse representation approaches, the proposed strategy significantly improves the representation quality.