Limited Memory Resources Research Articles

Enhancing IOT Security: Investigating Lightweight Encryption Algorithms for Resource-Constrained Devices

Abstract: In this work, we examine three LWCs: namely AES-128, SPECK, and ASCON, considering their presentation on resource-constrained IoT devices. With the large-scale deployment of IoT in all areas of industry, ensuring the security and efficiency of cryptographic implementations has become highly critical because typical IoT devices are usually characterized by limited memory, processing power, and energy resources. The study ascertains the memory usage, throughput, energy efficiency, and strength of security in the chosen algorithms by highly popular resource-constrained boards such as Arduino Nano and Arduino Micro. The benchmarks have shown that SPECK has the highest throughput, which is much efficient for constrained environments. One of the significant features of ASCON is that it has minimized memory usage and high energy efficiency. Therefore, it suits for long-term usage in IoT. AES-128 is secure but resource-intensive, so its usage may be irrelevant in very constrained environments. The results are of great value for IoT developers as a contribution to the selection of cryptographic algorithms based on the conditions and resource restrictions of the IoT devices involved so that optimized security solutions are achieved.

Benchmarking In-Sensor Machine Learning Computing: An Extension to the MLCommons-Tiny Suite

This paper proposes a new benchmark specifically designed for in-sensor digital machine learning computing to meet an ultra-low embedded memory requirement. With the exponential growth of edge devices, efficient local processing is essential to mitigate economic costs, latency, and privacy concerns associated with the centralized cloud processing. Emerging intelligent sensors equipped with computing assets to run neural network inferences and embedded in the same package, which hosts the sensing elements, present new challenges due to their limited memory resources and computational skills. This benchmark evaluates models trained with Quantization Aware Training (QAT) and compares their performance with Post-Training Quantization (PTQ) across three use cases: Human Activity Recognition (HAR) by means of the SHL dataset, Physical Activity Monitoring (PAM) by means of the PAMAP2 dataset, and superficial electromyography (sEMG) regression with the NINAPRO DB8 dataset. The results demonstrate the effectiveness of QAT over PTQ in most scenarios, highlighting the potential for deploying advanced AI models on highly resource-constrained sensors. The INT8 versions of the models always outperformed their FP32, regarding memory and latency reductions, except for the activations for CNN. The CNN model exhibited reduced memory usage and latency with respect to its Dense counterpart, allowing it to meet the stringent 8KiB data RAM and 32 KiB program RAM limits of the ISPU. The TCN model proved to be too large to fit within the memory constraints of the ISPU, primarily due to its greater capacity in terms of number of parameters, designed for processing more complex signals like EMG. This benchmark aims to guide the development of efficient AI solutions for In-Sensor Machine Learning Computing, fostering innovation in the field of Edge AI benchmarking, such as the one conducted by the MLCommons-Tiny working group.

A quantum-safe authentication scheme for IoT devices using homomorphic encryption and weak physical unclonable functions with no helper data

Physical Unclonable Functions (PUFs) are widely used to authenticate electronic devices because they take advantage of random variations in the manufacturing process that are unique to each device and cannot be cloned. Therefore, each device can be uniquely identified and counterfeit devices can be detected. Weak PUFs, which support a relatively small number of challenge-response pairs (CRPs), are simple and easy to construct. Device authentication with weak PUFs typically uses helper data to obfuscate and recover a cryptographic key that is then required by a cryptographic authentication scheme. However, these schemes are vulnerable to helper-data attacks and many of them do not protect conveniently the PUF responses, which are sensitive data, as well as are not resistant to attacks performed by quantum computers. This paper proposes an authentication scheme that avoids the aforementioned weaknesses by not using helper data, protecting the PUF response with a quantum-safe homomorphic encryption, and by using a two-server setup. Specifically, the CRYSTALS-Kyber public key cryptographic algorithm is used for its quantum resistance and suitability for resource-constrained Internet-of-Things (IoT) devices. The practicality of the proposal was tested on an ESP32 microcontroller using its internal SRAM as a SRAM PUF. For PUF responses of 512 bits, the encryption execution time ranges from 16.41 ms to 41.08 ms, depending on the desired level of security. In terms of memory, the device only needs to store between 800 and 1,568 bytes. This makes the solution post-quantum secure, lightweight and affordable for IoT devices with limited computing, memory, and power resources.

Optimizing Spatial Complexity of the Hard Decision Decoder Based on Hash and Syndrome Decoding (HSDec)

In this article, we propose an optimized version of the Hard Decision Decoder based on Hash and Syndrome Decoding (HSDec) decoder, named Reduced Memory Space of HSDec (RMS-HSDec), which uses less memory space. In this article, we aim to reduce the spatial complexity of the HSDec decoding algorithm while preserving its error correction capabilities. Our methodology involves allocating only the essential memory space for correctable error patterns and optimizing the hashing mechanism to effectively handle potential collisions. While maintaining the integrity of error correction, this new method guarantees memory reduction rates of over 96 % for the BCH(63, 39, 9) code and over 84 % for the QR(47, 24, 11) code compared to HSDec. Simulations were conducted to evaluate the performance of RMS-HSDec on various BCH and QR codes over AWGN and Rayleigh channels. The results demonstrated significant memory reduction rates and coding gains ranging from 0.8 dB to 2.8 dB over the AWGN channel and from 14 dB to 32 dB over the Rayleigh channel, confirming the robustness of the algorithm under different channel conditions. Comparative analyses showed that RMS-HSDec maintains competitive performance with existing decoders while offering effective error correction. These findings confirm the robustness of the RMS-HSDec algorithm under different channel conditions. Overall, the proposed decoder proves to be an effective solution, optimizing memory usage without compromising error correction capabilities, making it ideal for high-density data applications and environments with limited memory resources.

Three-dimensional grid-free sound source localization method based on deep learning

Sound source localization (SSL) technology is a popular method for identifying the locations of noise sources, which serves as a prerequisite for noise control. Deep learning, as a data-driven tool, shows broad perspectives in the field of SSL with its powerful nonlinear fitting ability. The existing deep learning-based SSL methods only provide a two-dimensional (2D) representation of the sound source location and cannot obtain the specific coordinates of the sound source in three-dimensional (3D) space. Although traditional beamforming methods can be directly generalized to 3D scenes in principle, they suffer from the limitations of insufficient vertical resolution and high computational cost. Therefore, a 3D grid-free SSL method (3DGF) informed by deep learning is suggested in this study to enhance the accuracy and computational efficiency of 3D localization. First, the number of data channels is compressed to respect limited memory resources during the training process. Subsequently, a dense convolutional neural network (DenseNet) model is utilized to obtain the 3D spatial coordinates of the sound source using the processed 3D beamforming map as input. Since the coordinates are continuous and are not constrained by the grid of the beamforming map, the grid-free strategy presents more accurate localization results. Then, the effects of the volume of training data and the compression ratio are analyzed, respectively, in simulation, and the localization performance with different signal-to-noise ratios (SNRs) is also tested. Finally, by comparing 3DGF with DAMAS, both simulation and experimental results demonstrate that 3DGF improves the accuracy and efficacy of 3D localization. Meanwhile, its satisfactory generalization ability and robustness against noise highlight its potential for practical applications.

Efficient Management of Containers for Software Defined Vehicles

Containerization technology, such as Docker, is gaining in popularity in newly established software-defined vehicle architectures (SDVA). However, executing those containers can quickly become computationally expensive in constrained environments, given the limited CPU, memory, and energy resources in the Electric Control Units (ECU) of SDVA. Consequently, the efficient management of these containers is crucial for enabling the on-demand usage of the applications in the vehicle based on the available resources while considering several constraints and priorities, including failure tolerance, security, safety, and comfort. In this paper, we propose a dynamic software container management approach for constrained environments such as embedded devices/ECUs in SDVA within smart cars. To address the conflicting objectives and constraints within the vehicle, we design a novel search-based approach based on multi-objective optimization. This approach facilitates the allocation, movement, or suspension of containers between ECUs in the cluster. Collaborating with our industry partner, Ford Motor Company, we evaluate our approach using different real-world software-defined scenarios. These scenarios involve using heterogeneous clusters of ECU devices in vehicles based on real-world software containers and use-case studies from the automotive industry. The experimental results demonstrate that our scheduler outperforms existing scheduling algorithms, including the default Docker scheduler -Spread- commonly used in automotive applications. Our proposed scheduler exhibits superior performance in terms of energy and resource cost efficiency. Specifically, it achieves a 35% reduction in energy consumption in power-saving mode compared to the scheduler employed by Ford Motor Company. Additionally, our scheduler effectively distributes workload among the ECUs in the cluster, minimizing resource usage, and dynamically adjusts to the real-time requirements and constraints of the car environment. This work will serve as a fundamental building block in the automotive industry to efficiently manage software containers in smart vehicles considering constraints and priorities in the real world.

Design of an optimized sensor fault identification within the limited computing, memory, and energy capabilities

The main goal of this work is to design an optimized sensor-fault identification and diagnostic system for the Internet of Things (IoT) and Cyber-Physical Systems (CPS). The challenge is to accomplish this task within the sensors’ limited computing, memory, and energy capabilities. More importantly, identifying errors is time-sensitive, even though the diagnosis does not have to be made quickly. This project aims to provide an enhanced sensor-fault detection and diagnostic system for the IoT and CPS with constrained energy, memory, and computation resources. The system’s goals are to promptly detect defects, lessen the computing burden on sensors, and enhance the recall and accuracy of fault detection. This study used a hybrid approach that combined principal component analysis, autoencoder, and gated recurrent unit to create an optimal sensor-fault detection and diagnostic system. There were 1001 sensor readings in the dataset; 112 were defective, while the remaining 888 were normal. The investigation showed that the suggested method, which detected faults with an accuracy of 95% and recall of 92%, achieved high accuracy and recall in recognizing defects in IoT and CPS. With significantly shorter processing times, the system’s potential to reduce computational strain on sensors was also proved. The findings of this study indicate that the suggested optimized sensor-fault detection and diagnosis system successfully detects faults in IoT and CPS with limited computation, memory, and energy resources. The system’s ability to reduce the computing burden on sensors while improving accuracy and recall makes it an appealing choice for industrial and commercial applications.

DESIGN: Online Device Selection and Edge Association for Federated Synergy Learning-enabled AIoT

The Artificial Intelligence of Things (AIoT) is an emerging technology that enables numerous AIoT devices to participate in big data analytics and machine learning (ML) model training, providing various customized intelligent services for industry manufacturing. Federated Learning (FL) empowers AIoT applications with privacy-preserving distributed model training without sharing raw data. However, due to IoT devices’ limited computing and memory resources, existing FL approaches for AIoT applications cannot support efficient large-scale model training. Federated synergy learning (FSyL) is a promising collaborative paradigm that alleviates the computation and communication overhead on resource-constrained AIoT devices via offloading part of the ML model to the edge server for end-to-edge collaborative training. Existing FSyL works neither efficiently address the inter-round device selection to improve model diversity nor determine the intra-round edge association to reduce the training cost, which hinders the applications of FSyL-enable AIoT. Motivated by this issue, this paper first investigates the bottlenecks of executing FSyL in AIoT. It builds an optimization model of joint inter-round device selection and intra-round edge association for balancing model diversity and training cost. To tackle the intractable coupling problem, we present a framework named Online DE vice S elect I on and Ed G e Associatio N for Cost-Diversity Trade-offs FSyL (DESIGN). First, the edge association subproblem is extracted from the original problem, and game theory determines the optimal association decision for an arbitrary device selection. Then, based on the optimal association decision, device selection is modeled as a combinatorial multi-armed bandit (CMAB) problem. Finally, we propose an online mechanism to obtain joint device selection and edge association decisions. The performance of DESIGN is theoretically analyzed and experimentally evaluated on real-world datasets. The results show that DESIGN can achieve up to \(84.3\%\) in cost-saving with an accuracy improvement of \(23.6\%\) compared with the state-of-the-art.

Representation and computation in visual working memory.

The ability to sustain internal representations of the sensory environment beyond immediate perception is a fundamental requirement of cognitive processing. In recent years, debates regarding the capacity and fidelity of the working memory (WM) system have advanced our understanding of the nature of these representations. In particular, there is growing recognition that WM representations are not merely imperfect copies of a perceived object or event. New experimental tools have revealed that observers possess richer information about the uncertainty in their memories and take advantage of environmental regularities to use limited memory resources optimally. Meanwhile, computational models of visuospatial WM formulated at different levels of implementation have converged on common principles relating capacity to variability and uncertainty. Here we review recent research on human WM from a computational perspective, including the neural mechanisms that support it.

A Novel Attention‐Based Layer Pruning Approach for Low‐Complexity Convolutional Neural Networks

Deep learning (DL) has been very successful for classifying images, detecting targets, and segmenting regions in high‐resolution images such as whole slide histopathology images. However, analysis of such high‐resolution images requires very high DL complexity. Several AI optimization techniques have been recently proposed that aim at reducing the complexity of deep neural networks and hence expedite their execution and eventually allow the use of low‐power, low‐cost computing devices with limited computation and memory resources. These methods include parameter pruning and sharing, quantization, knowledge distillation, low‐rank approximation, and resource efficient architectures. Rather than pruning network structures including filters, layers, and blocks of layers based on a manual selection of a significance metric such as l1‐norm and l2‐norm of the filter kernels, novel highly efficient AI‐driven DL optimization algorithms using variations of the squeeze and excitation in order to prune filters and layers of deep models such as VGG‐16 as well as eliminate filters and blocks of residual networks such as ResNet‐56 are introduced. The proposed techniques achieve significantly higher reduction in the number of learning parameters, the number of floating point operations, and memory space as compared to the‐state‐of‐the‐art methods.

Quick UDP Internet Connections and Transmission Control Protocol in unsafe networks: A comparative analysis

AbstractSecure data transmission and efficient network performance are both key aspects of the modern Internet. Traditionally, Transport Layer Security (TLS)/Transmission Control Protocol (TCP) has been used for reliable and secure networking communications. In the past decade, Quick User Datagram Protocol (UDP) Internet Connections QUIC has been designed and implemented on UDP, attempting to improve security and efficiency of Internet traffic. Real‐world platform investigations are carried out in this paper to evaluate TLS/TCP and QUIC/UDP in maintaining communication, security and efficiency under three different types of popular cyber‐attacks. A set of interesting findings, including delay, loss, server CPU utilisation and server memory usage are presented to provide a comprehensive understanding of the two protocol stacks in performing malicious traffic. More specifically, in terms of the efficiency in achieving short delays and low packet loss rates with limited CPU and memory resources, QUIC/UDP performs better under Denial of Service attacks but TLS/TCP overtakes QUIC/UDP when handling MitM attacks. In terms of security, the implementation of TCP tends to be more secure than QUIC, but QUIC traffic patterns are harder to learn using machine learning methods. We hope that these insights will be informative in protocol selection for future networks and applications, as well as shedding light on the further development of the two protocol stacks.

Internet of Things intrusion detection: Research and practice of NSENet and LSTM fusion models

To address the problems of complex environment, limited device computational resources and limited memory resources in the existing IoT, SELSTM, an intrusion detection system composed of NSENet and LSTM fusion based on SENet, is investigated. The NSENet part of the SELSTM system is based on the squeeze-and-excitation network (SENet). The lightweight computational modules NonLocal, SKConv and inverted residuals are fused into SE blocks, and self-attention of Nonlocal is used to improve the local receptive field of feature extraction. The channel attention and spatial attention of each part of the data are strengthened by the use of SKConv To enhance the adaptive convolution ability of the model and ensure the completeness of the information, the properties of the inverted residual structure are used to ensure that the gradient of the model decreases steadily without gradient explosion or disappearance. For the problem of data imbalance, the dataset is randomly resampled using the weight resampling technique to improve the balance of the dataset to ensure that the final detection effect of the model is more effective and generalized, while the data flow is divided into two parts for processing, and the model parameters are optimized using the model gradient optimizer consisting of the optimizer Lion and the optimization function Lookahead. The model extracts the spatial and temporal features of the data through multidimensional extraction to ensure the completeness of the data feature information in multiple dimensions, thus obtaining better detection results. The results of the experiments comparing the SELSTM model with other models on the intrusion dataset show that the intrusion detection model has a higher detection precision and accuracy than the traditional deep learning intrusion detection model, which indicates that the SELSTM has better detection performance properties and better practicality and effectiveness on IoT devices.

Accelerating page loads via streamlining JavaScript engine for distributed learning

Distributed learning based on JavaScript-based frontends is typically implemented at the endpoint to maximize performance. Yet, JavaScript-based frontends often experience suboptimal performance. To reconcile these disparities in performance between EDGE and endpoint deployments, strategic optimization is essential, particularly for preserving privacy in distributed learning. Real-time streaming optimizations are imperative to align the performance of disparate components for smooth integration. The reliance on JavaScript for various web functionalities can lead to increased resource consumption and slower page loads. Thus, we introduce a streamlined JavaScript engine designed to optimize structural patterns in JavaScript code, with three key enhancements. Firstly, we reduce the computational burden of the JavaScript engine necessary for setting up the browser's runtime environment. Secondly, we refine the parsing process for specific code patterns, boosting the efficiency of our lightweight engine. Thirdly, we streamline the Inter-Process Communication (IPC) to maintain high performance, even with limited memory resources. Our evaluations demonstrate that our approach reduces the median Total Computation Time (TCT) by 45.2%, and surpasses existing leading solutions, Siploader and Prepack, with improvements ranging from 1.13× to 1.39×.

Adaptive forgetting speed in working memory.

Working memory is known to be capacity-limited and is therefore selective not only for what it encodes but also what it forgets. Explicit forgetting cues can be used effectively to free up capacity, but it is not clear how working memory adaptively forgets in the absence of explicit cues. An important implicit cue that may tune forgetting in working memory is the passage of time. When information becomes irrelevant more quickly, working memory should also forget information more quickly. In three delayed-estimation experiments, we systematically manipulated how probing probability changed as time passed on after encoding an item (i.e., the "probing hazard"). In some blocks, probing hazard decreased after encoding an item, requiring participants to only briefly retain the memory item. In other blocks, the probing hazard increased or stayed flat, as the retention interval was lengthened. In line with our hypothesis, we found that participants adapted their forgetting rate to the probing dynamics of the working memory task. When the memory item quickly became irrelevant ("decreasing" probing hazard), forgetting rate was higher than in blocks where probing hazard increased or stayed flat. The time course of these adaptations in forgetting implies a fast and flexible mechanism. Interestingly, participants could not explicitly report the order of conditions, suggesting forgetting is implicitly sped up. These findings suggest that implicit adaptations to the temporal structure of our environment tune forgetting speed in working memory, possibly contributing to the flexible allocation of limited working memory resources.

Eye movement intervention facilitates concurrent perception and memory processing.

A widely used psychotherapeutic treatment for post-traumatic stress disorder (PTSD) involves performing bilateral eye movement (EM) during trauma memory retrieval. However, how this treatment-described as eye movement desensitization and reprocessing (EMDR)-alleviates trauma-related symptoms is unclear. While conventional theories suggest that bilateral EM interferes with concurrently retrieved trauma memories by taxing the limited working memory resources, here, we propose that bilateral EM actually facilitates information processing. In two EEG experiments, we replicated the bilateral EM procedure of EMDR, having participants engaging in continuous bilateral EM or receiving bilateral sensory stimulation (BS) as a control while retrieving short- or long-term memory. During EM or BS, we presented bystander images or memory cues to probe neural representations of perceptual and memory information. Multivariate pattern analysis of the EEG signals revealed that bilateral EM enhanced neural representations of simultaneously processed perceptual and memory information. This enhancement was accompanied by heightened visual responses and increased neural excitability in the occipital region. Furthermore, bilateral EM increased information transmission from the occipital to the frontoparietal region, indicating facilitated information transition from low-level perceptual representation to high-level memory representation. These findings argue for theories that emphasize information facilitation rather than disruption in the EMDR treatment.

Binary Graph Convolutional Network With Capacity Exploration.

The current success of Graph Neural Networks (GNNs) usually relies on loading the entire attributed graph for processing, which may not be satisfied with limited memory resources, especially when the attributed graph is large. This paper pioneers to propose a Binary Graph Convolutional Network (Bi-GCN), which binarizes both the network parameters and input node attributes and exploits binary operations instead of floating-point matrix multiplications for network compression and acceleration. Meanwhile, we also propose a new gradient approximation based back-propagation method to properly train our Bi-GCN. According to the theoretical analysis, our Bi-GCN can reduce the memory consumption by an average of ∼31x for both the network parameters and input data, and accelerate the inference speed by an average of ∼51x, on three citation networks, i.e., Cora, PubMed, and CiteSeer. Besides, we introduce a general approach to generalize our binarization method to other variants of GNNs, and achieve similar efficiencies. Although the proposed Bi-GCN and Bi-GNNs are simple yet efficient, these compressed networks may also possess a potential capacity problem, i.e., they may not have enough storage capacity to learn adequate representations for specific tasks. To tackle this capacity problem, an Entropy Cover Hypothesis is proposed to predict the lower bound of the width of Bi-GNN hidden layers. Extensive experiments have demonstrated that our Bi-GCN and Bi-GNNs can give comparable performances to the corresponding full-precision baselines on seven node classification datasets and verified the effectiveness of our Entropy Cover Hypothesis for solving the capacity problem.

A comparative study of large-scale and lightweight convolutional neural networks for ImageNet classification

In the field of convolution neural networks (CNNs), many impressive architectures have been published in recent years. These can be roughly divided into two groups: large-scale models and lightweight models. These large models are characterized by many trainable weights and complex network structures, offering them strong effectiveness in various computer vision tasks, and have become essential components of many modern visual recognition systems. These lightweight CNNs are designed to maintain high performance with limited memory and computational resources. They are highly efficient in terms of inference time and resource utilization, so that particularly suitable for mobile and edge computing devices. This work focuses on some prominent models based on the ImageNet database and explores the reasons for their frameworks success. By analyzing these models, a trend could be identified in the development of CNN models: Reasonably increasing the scale of the model and utilizing suitable frameworks can both improve accuracy and efficiency.

Spatial Transformation Accelerator with Parallel Data Access Scheme for Sample Reconstruction

Spatial image transformation is a commonly used component in many image processing pipelines. It enables the correction of optical distortions, image registration onto a common reference plane, electronic image stabilisation, digital zoom, video mosaicking, etc. With the growing tendency to embed image processing in low-power devices, attaining an efficient transformation solution becomes increasingly decisive. Furthermore, interpolation is the key operation in achieving the high quality of the transformed data from the original data. Fortunately, different implementations have already seen several efficiency improvements in recent years. However, interpolation relies on sampling a set of neighbouring points from memory, which has yet to be addressed efficiently for smaller computational platforms with limited memory resources. In this work, we derive a generic mathematical model and circuit design principles for the spatial transformation accelerator design for N-dimensional data. Furthermore, we present an efficient simultaneous access scheme for high-quality signal reconstruction. Finally, the introduced ideas are verified in field programmable gate arrays using one-dimensional and two-dimensional data transformation use cases. The presented solution is able to transform images with sizes ranging from 256 × 256 to 8192 × 8192 and achieves a transfer rate of 275 frames per second with 512 × 512 images.

CrossPrune: Cooperative pruning for camera–LiDAR fused perception models of autonomous driving

Deep neural network pruning is effective in enabling high-performance perception models to be deployed on autonomous driving platforms with limited computation and memory resources. With their rapid development, state-of-the-art autonomous driving perception (ADP) models have advocated the use of multimodal sensors to extract diverse feature categories. However, existing pruning studies in the ADP area focus on single-modal models and neglect multimodal models. Compared with conventional pruning, multimodal pruning presents a new type of redundancy, namely modal-wise redundancy that is caused by multimodal branches extracting similar perception information. When a specific type of information is extracted by more than one modal branch, although this extraction is deemed essential from an individual single-modal standpoint, it becomes redundant when viewed from a modal-wise perspective. Therefore, modal branches must be handled cooperatively to eliminate modal-wise redundancy while concurrently preserving the original perception accuracy. Building on this, we propose CrossPrune, a modal cooperative pruning framework designed for camera–LiDAR fused perception in autonomous driving. The primary objective of CrossPrune is to effectively eliminate redundancy and achieve nondestructive pruning for multimodal ADP models. This was accomplished by approaching the problem as a multi-objective optimization task, encompassing both weight pruning and the restriction of feature distortions caused by pruning. Experiments conducted on the nuScenes and KITTI datasets demonstrated that CrossPrune attained superior pruning ratios while minimizing accuracy loss, surpassing the performance of the baselines. The key results indicated that the proposed CrossPrune achieved relative improvements of 9.6% in mAP and 11.5% in NDS under 89.8% pruning sparsity.

A lightweight CNN for multi-source infrared ship detection from unmanned marine vehicles

Infrared ship detection is of great significance due to its broad applicability in maritime surveillance, traffic safety and security. Multiple infrared sensors with different spectral sensitivity provide enhanced sensing capabilities, facilitating ship detection in complex environments. Nevertheless, current researches lack discussion and exploration of infrared imagers in different spectral ranges for marine objects detection. Furthermore, for unmanned marine vehicles (UMVs), e.g., unmanned surface vehicles (USVs) and unmanned ship (USs), detection and perception are usually performed in embedded devices with limited memory and computation resource, which makes traditional convolutional neural network (CNN)-based detection methods struggle to leverage their advantages. Aimed at the task of sea surface object detection on USVs, this paper provides lightweight CNNs with high inference speed that can be deployed on embedded devices. It also discusses the advantages and disadvantages of using different sensors in marine object detection, providing a reference for the perception and decision-making modules of USVs. The proposed method can detect ships in short-wave infrared (SWIR), long-wave infrared (LWIR) and fused images with high-performance and high-inference speed on an embedded device. Specifically, the backbone is built from bottleneck depth-separable convolution with residuals. Generating redundant feature maps by using cheap linear operation in neck and head networks. The learning and representation capacities of the network are promoted by introducing the channel and spatial attention, redesigning the sizes of anchor boxes. Comparative experiments are conducted on the infrared ship dataset that we have released which contains SWIR, LWIR and the fused images. The results indicate that the proposed method can achieve high accuracy but with fewer parameters, and the inference speed is nearly 60 frames per second (FPS) on an embedded device.