ABSTRACT Blockchain introduced secure, peer‐to‐peer transactions in decentralized environments without reliance on central authorities. However, attackers continue to exploit blockchain projects. Over the past decade, reported blockchain exploit losses have exceeded $24 billion. Although existing research analyzes blockchain vulnerabilities and real‐world incidents, many studies focus on specific vulnerability classes or smart contract issues only. Therefore, we present an empirical study of major blockchain exploits from 2020 to 2025. Based on detailed analysis of these exploits, we introduce a practitioner‐oriented taxonomy that organizes the dominant root causes observed in the analyzed exploits across four domains: Human Risks, Off‐Chain Infrastructure Vulnerabilities, Private Key Compromises, and Smart Contract Vulnerabilities. Across these categories, we present case studies, code illustrations where applicable, and mitigation strategies. Moreover, we evaluate five vulnerability‐detection tools spanning four paradigms, namely static analysis ( Slither , 4naly3er ), symbolic execution ( Mythril ), bounded/formal verification ( SMTChecker ), and LLM‐assisted analysis ( GPTScan ), on a curated test suite of vulnerable and secure contracts. Our results show detection gaps under the evaluated setting: Slither, Mythril, and 4naly3er all scored recall below 40%. SMTChecker and GPTScan were evaluated separately because of their narrower target scopes. SMTChecker detected only arithmetic issues, while GPTScan produced no benchmark‐relevant detections. The study highlights the high average loss associated with smart contract access‐control failures, the need for post‐deployment monitoring beyond one‐time audits, and risks related to Off‐Chain Infrastructure Vulnerabilities. By analyzing exploits and offering mitigations, we provide a structured reference for identifying, prioritizing, and mitigating major blockchain security risks.
ABSTRACT The regulation has required regulatory‐compliant steganalysis to securely deliver medical images in telemedicine with a high level of accuracy in detection, resistance to adversarial attacks, and clinical interpretability. Although recent LSTM‐CNN hybrids offer parameter efficiency, they do not provide pixel‐level explainability, global spatial modeling, and can be compared to adaptive attacks that are essential obstacles to the usage of FDA (Food and Drug Administration). We present SA‐Net‐Stego‐Medical a new four‐stage model that is trained on clinically realistic datasets: ChestMNIST chest x‐rays and PathMNIST pathology slides. SA‐Net (Self‐Attention Network) replaces sequential LSTM (Long‐Short Term Memory) with multi‐head self‐attention, adding CBAM (Convolutional Block Attention Module) attention to SRM (Spatial Rich Model) residuals, and using the Grad‐CAM++ explainability, outperforms baselines significantly in both datasets and also shows excellent cross‐domain generalization. Grad‐CAM++ heatmaps accurately identify steganographic features to embedding‐prone anatomic structures (lung parenchyma, gland tissue) and get high plausibility scores with board‐certified radiologists. Adversarial training is resistant to FGSM (Fast Gradient Sign Method) and PGD (Projected Gradient Descent) attacks, and is significantly more resistant than baseline methods. There is the first production‐ready platform of secure, explainable medical imaging security in adversarial telemedicine setting with compact parameterization, real‐time inference, and full FDA 21 CFR (Code of Federal Regulations) Part 11 transparency.
ABSTRACT DICOM medical images contain sensitive diagnostic information, which is exposed to unauthorized access during transmission in open networks and thus needs encryption with strong security and low computational overhead. Existing DNA‐based image encryption schemes are limited by chaotic randomness, computational efficiency, and cryptanalytic attack. To overcome these limitations, a selective Region of Interest (ROI)‐based encryption framework is proposed, where diagnostically significant regions are encrypted to reduce the computational cost associated with full‐image processing. The framework integrates a 3D hybrid chaotic random number generator (RNG), SHA‐256‐based key generation, and dynamic DNA rule mapping. The hybrid RNG combines multiple nonlinear chaotic systems to achieve a wider chaotic range, more uniform output, and improved randomness under finite‐precision computation; its dynamical characteristics are validated through Lyapunov spectra, bifurcation diagrams, periodic‐window analysis, and NIST statistical tests, with positive Lyapunov exponents confirming chaotic behavior and statistical tests demonstrating sequence randomness. SHA‐256 derives image‐dependent initial conditions to enhance key and plaintext sensitivity and resistance to brute‐force and known‐plaintext attacks. Dynamic DNA rules are adaptively selected from chaotic sequences to increase mapping uncertainty and strengthen confusion and diffusion. Experimental results demonstrate high entropy, low inter‐pixel correlation, strong differential resistance, and a throughput of 1.192 Mbps for 512 × 512 images, demonstrating an effective security–efficiency trade‐off.
ABSTRACT The security of symmetric‐key block ciphers depends critically on the strength and structural unpredictability of their substitution boxes (S‐boxes). Existing constructions rely on fixed mathematical structures, such as finite field inversion, chaotic maps, or elliptic curves that are entirely independent of the encrypting user's identity, creating a fundamental separation between authentication and encryption layers. This article presents a formally analyzed framework that derives all cryptographic parameters of a block cipher directly and systematically from the five biometric feature classes of a fingerprint image: minutiae points, ridge orientation fields, singular points (cores and deltas), ridge frequency maps, and local block entropy. Each feature class is mapped, through a novel algebraic procedure, to a distinct cipher parameter: (i) minutiae coordinates and angles seed a Fisher‐Yates permutation that whitens the output of the S‐box permutation layer; (ii) the ridge orientation field is binarized and corrected to generate an invertible matrix for the affine diffusion transform; (iii) singular point topology determines the biometric‐adaptive round count ; (iv) the mean ridge frequency selects the irreducible polynomial from a set of four valid polynomials; and (v) the block entropy map calibrates the per‐round diffusion strength parameter . Three multi‐round block encryption algorithms of increasing complexity are built upon this five‐parameter biometric cipher core, evaluated with a full Python reference implementation that reproduces every reported figure. The primary scheme (Algorithm II) achieves a plaintext avalanche effect of and a key sensitivity of (16 rounds, trials), reaching within of the ideal from round 1. The fingerprint S‐boxes exhibit nonlinearity values in the range with mean 103.8, differential uniformity , SAC mean , BIC‐NL average , and algebraic degree 7 on all eight output bits. A critical cross‐user security property is demonstrated: encrypting identical plaintext with an identical conventional key but different fingerprint templates produces ciphertexts whose pairwise Hamming distance averages (range – per trial, – averaged over 30 trials per pair), confirming that the biometric parameter functions as a genuine second key factor. A parameter‐stability experiment under simulated repeated‐acquisition noise (translation, rotation, minutiae dropout and spurious detections) further shows that the coarse, quantized cipher parameters (round count, field polynomial, diffusion strength) remain stable, while the fine‐grained minutiae‐derived permutation does not—a scope limitation discussed openly alongside the reliance on synthetic ISO/IEC 19794‐2‐format templates.
ABSTRACT The rapid growth of the Internet of Things (IoT) creates concern around multilayer cyber‐attacks that exploit interactions between physical, network, and application layers. Traditional Intrusion Detection Systems (IDS) often utilize outdated datasets, high dimensional features, or computationally heavy deep learning models, which are not practical for low‐power IoT environments. This study extends previous research on the Semi‐Automated Intrusion Detection System (SAIDS) framework by evaluating its performance using a hybrid dataset constructed from both simulated and real‐world IoT attack scenarios. The dataset combines Denial of Service (UDP flood) traffic generated using the Cooja simulator with a real‐time ARP‐based Man‐in‐the‐Middle (MITM) attack executed on a smart home IoT testbed. A cross‐layer feature analysis was conducted to identify behavioral features shared across the selected attack types. From an initial set of 35 features, 21 common features were identified, and a reduced subset of seven lightweight features was selected using multiple feature selection and weighting methods supported by a semi‐automated analysis process. Five machine learning models, KNN, DT, ANN, RF, and NB, were evaluated using both binary and multiclass classification approaches. The results show that DT, ANN, and KNN achieved strong performance, with accuracy ranging from 84% to 100%, demonstrating the effectiveness of the reduced feature set in detecting representative multilayer attack scenarios involving DoS and MITM. In contrast to existing studies that rely primarily on benchmark datasets or high‐complexity models, this work demonstrates that lightweight feature subsets can maintain strong detection performance across heterogeneous IoT environments. The findings highlight the practical applicability of the SAIDS framework for real‐world IoT intrusion detection, particularly in scenarios where computational efficiency and adaptability are required.
