Figure 1.
Blockchain process.
Table 1.
Overview of Major Consensus Algorithms in Blockchain Systems
| Consensus Mechanism | Mechanism | Security Strengths | Weaknesses | Energy Efficiency | Scalability |
|---|---|---|---|---|---|
| Proof of Work (PoW) [12] | Miners solve cryptographic puzzles | High security, prevents Sybil attacks | Energy-intensive, slow transactions | Very Low | Poor |
| Proof of Stake (PoS) [13] | Validators are chosen based on the number of coins staked | Resistant to Sybil attacks through staking requirements | Susceptible to "nothing at stake" attacks | High | Good |
| Delegated Proof of Stake (DPoS) [14] | Selected nodes validate transactions on behalf of others | Fast finality and reduced attack surface with fewer nodes | Potential centralization risks | High | Excellent |
| Practical Byzantine Fault Tolerance (PBFT) [15] | Nodes reach consensus through a series of voting rounds | Tolerant to malicious nodes | Does not scale well for large networks | High | Poor to Moderate |
Table 2.
Deep Learning Architectures to Enhance Blockchain Security
| Deep Learning Model | Application in Blockchain Security | Strengths | Limitations |
|---|---|---|---|
| Convolutional Neural Networks (CNNs) [17] | Malware detection, smart contract security analysis | Extracts spatial patterns, high accuracy | Requires large labeled datasets |
| Recurrent Neural Networks (RNNs) [18] | Fraud detection, anomaly detection in transactions | Captures sequential dependencies | Training complexity, vanishing gradient problem |
| Long Short–Term Memory (LSTMs) [19] | Intrusion detection, network traffic analysis | Overcomes short-term memory issues in RNNs | Requires high computational power |
| Autoencoders [20] | Detecting anomalies in blockchain transactions | Learns compressed representations, detects outliers | May not work well for real-time detection |
| Graph Neural Networks (GNNs) [21] | Analyzing blockchain transaction networks | Captures complex entity relationships | High training cost, requires large-scale blockchain data |
Figure 2.
Distribution of the surveyed papers by publication year and dominant deep-learning family.
Figure 3.
Master Taxonomy: Threats, Integration Locus, DL Families, Data Modalities, and Deployment Settings
Figure 4.
Blockchain Security Vulnerabilities
Table 3.
Various Deep Learning Approaches Have Been Applied Blockchain Fraud Detection
| Criteria | Blockchain Focus | Blockchain Integration | Main Technique | Dataset | Performance | Key Contributions | Weaknesses/Challenges |
|---|---|---|---|---|---|---|---|
| Lakhan et al. (2022) [26] | Serverless Blockchain (SBETS) | Hybrid | Deep Graph Convolutional Network (GCN) | Simulated ITS dataset | 10% cost reduction, 50% fraud detection gain | Serverless task scheduling and secure ITS fraud detection with GCN | Simulation-based, lacks real-world testing |
| Roy et al. (2022) [23] | Ethereum | Off-chain | LSTM + Dense Neural Network with Information Gain for feature selection | Ethereum transaction dataset | Significant improvement over baseline models | Introduces a reliable Ethereum fraud detection model using LSTM; highlights need for quality datasets; feature selection reduces model complexity | Lack of public dataset limits reproducibility; focuses only on binary classification; limited explanation of model decisions |
| Ghosh et al. (2023) [24] | Bitcoin | Off-chain | Ensemble of MLP, FNN, and Attention-based LSTM | Bitcoin transaction data | Accuracy: 99.62%; Precision & Recall >99% | Combines multiple deep learning models to achieve superior accuracy and robustness in fraud detection | Only focused on Bitcoin network; lacks model interpretability; dataset specifics are missing |
| Olusegun and Yang (2024) [25] | Ethereum | Off-chain | Ensemble of deep learning models (MLP, LSTM, CNN, CLSTM, TabPFN) with SHAP-based feature selection and explainable AI | Ethereum transaction dataset | Accuracy: 99.2% | Incorporates explainable AI (SHAP) and transformer-based TabPFN to provide both high accuracy and transparency in fraud detection | Complexity and potential overfitting from ensemble models; limited real-world deployment discussion |
| Airlangga (2024) [31] | Open Metaverse blockchain | Off-chain | MLP, CNN, LSTM | metaverse transactions | MLP achieved lowest MSE | Comparative DL approach to metaverse fraud detection | Offline only; lacks real-time capability |
| Taher et al. (2024) [27] | Ethereum | Off-chain | Ensemble Learning + XAI (LIME) | Ethereum transaction dataset | 99% accuracy | Ensemble + XAI for interpretable fraud detection | Scalability and real-time deployment not evaluated |
| Baabdullah et al. (2024) [29] | Credit card blockchain-integrated | Hybrid | Federated Learning + Blockchain, LSTM, CNN, RF | Federated dataset from 3 banks | Improved accuracy and privacy | Privacy-preserving fraud detection via FL and blockchain | Trade-off between latency and privacy; dataset access |
| Asiri and Somasundaram (2025) [28] | Bitcoin | Off-chain | Graph Convolutional Network (GCN) | Elliptic Bitcoin Dataset | 98.5% accuracy, AUC 0.9444, RMSE 0.1123 | GCN applied to illicit Bitcoin detection | Dataset partially labeled; possible generalization issues |
| Gu and Dib (2025) [30] | Ethereum | Off-chain | Ensemble (RF, SVM, XGBoost) | Ethereum transaction dataset | >98% on all metrics; 0.13s inference time | Real-time, high-performance ensemble fraud detection model | Class imbalance in Ethereum transaction data |
| Sun et al. (2025) [32] | Ethereum | Off-chain | Hybrid LM+Graph | MulDiGraph, B4E, SPN | +10–20% F1 over SOTA across all three datasets | First to fuse transaction semantics (LM) with similarity and structure under joint optimization; releases dataset | Complexity & compute; vocabulary/design choices; interpretability of fused views |
| Kang and Hwang (2025) [33] | Ethereum | Off-chain | Transformer with Random Traversal over the transaction graph + deep graph features | Large-scale Ethereum transaction networks | Higher AUC than baselines on large networks | Scalable traversal strategy to form sequences for a transformer while preserving graph context | Potential training cost; explaining decisions may be non-trivial |
| Gomes and Ferreira (2025) | Bitcoin | Off-chain | Dimensionality reduction + ML (PCA/UMAP + FS) with multiple classifiers; XGBoost best | Elliptic Bitcoin dataset | 3-class: XGBoost around 0.926 acc; 2-class performs higher; | Clean pipeline highlighting explainability via reduced feature sets; comprehensive baseline sweep | Imbalance & label scarcity |
Figure 5.
An example of smart contracts [35]
Table 4.
Various Deep Learning Approaches Have Been Applied Blockchain smart contract
| Criteria | Blockchain Focus | Blockchain Integration | Main Technique | Dataset | Performance | Key Contributions | Weaknesses/Challenges |
|---|---|---|---|---|---|---|---|
| Li et al. (2022) [38] | Ethereum, VNTChain | Off-chain | Deep and Cross Network (Link-DC model) combining contract graphs and expert features | Smart contracts from Ethereum and VNTChain | Reentrancy: 94.37%, Timestamp Dependency: 92.11%, Infinite Loop: 85.29% | Fuses expert patterns with constructed graph features for richer context | Relies on manually defined expert patterns; possible overfitting to known patterns |
| Li et al. (2023) [37] | Ethereum | Off-chain | Multimodal feature fusion using GCN and Bi-LSTM | Ethereum contracts | Reentrancy: 85.73%, Timestamp Dependency: 85.41%, | Combines static and dynamic features; uses deep learning for comprehensive multimodal analysis | May have limited generalization to unseen vulnerabilities; performance may vary with dataset balance |
| Han et al. (2023) [41] | Ethereum | Off-chain | Syntactic and Semantic Feature Fusion using TextCNN and GNN | Ethereum Smart Contracts (AST + CFG) | Precision: 96%, Recall: 90% | Fusion of AST and CFG features for better vulnerability detection | Limited to five vulnerability types, may not scale |
| Deng et al. (2023) [44] | Ethereum | Off-chain | Deep Learning with Multimodal Decision Fusion | Public Smart Contract Dataset | Accuracy: 94.8%, AUC: 0.886 | Fuses multi-modal information (SC, OP, CFG) via DL | High computational overhead with multiple modalities |
| Hwang et al. (2024) [40] | Ethereum | Off-chain | Compiler-Guided Generation with Monte Carlo Tree Search | Generated via CGGNet (3.37M contracts) | 3,369,397 contracts generated | First smart contract-focused augmentation using compiler-guided learning | Depends on compiler and MCTS configuration quality |
| Osei et al. (2024) [39] | Ethereum | Off-chain | Wide & Deep Neural Network (WIDENNET) using opcode vectors | Real-world smart contracts | Accuracy: 83.07%, Precision: 83.13% | Introduces wide and deep learning model for combining general and specific vulnerability patterns | Limited to detecting reentrancy and timestamp dependency; lower performance compared to other models |
| Li et al. (2024) [42] | Ethereum (DEXs) | Off-chain | Graph-based Deep Learning using GCN on AST Dependencies | 46 DEX Projects Dataset (5,671 | Precision: 92.24%, F1-score: 94.25% | Detection of state derailment in DEXs using dependency-aware GCN | Focus only on DEX-related state issues |
| Kasula et al. (2024) [43] | Ethereum | Off-chain | Graph-based Deep Learning Approach | Ethereum explorer | N/R | Integrates structural and semantic contract features | Limited methodological detail; |
| Chen et al. (2024) [35] | Ethereum | Off-chain | Contrastive Learning with Transformer-based Architecture | 40K+ Real-world Contracts | F1-score improved by 9.73% to 39.99% | Applies contrastive learning for improved contract context understanding | Requires pair-wise contract samples and correlation labels |
| Wang et al. (2025) [45] | Ethereum | Off-chain | Multimodal, multi-scale Transformer fusion over code tokens + AST/CFG | Public Ethereum contracts | N/R | Shows complementarity of text + structure; scalable fusion design | Potential compute overhead for graph + transformer; generalization beyond tested vulns not detailed |
| Balci et al. (2025) [46] | Ethereum | Off-chain | Transformer over EVM bytecode sequences | ~16,4xx verified 2022 contracts | N/R | Works without source code; strong scalability claims; contemporary, large verified set | Potential label noise from auto-annotation; need reproducible release of data/code |
| Choi et al. (2025) | Ethereum | Off-chain | CodeBERT (text) + CFG/AST/opcode graphs, Sent2Vec + centrality | Not mentioned | Acc 86.70%, Prec 85.24%, Rec 84.87%, F1 84.46% | Clear text–structure fusion recipe; transparent metrics; | Moderate F1 vs newer multimodal systems; dependency on LLM prompting; compute for graph extraction; |
Table 5.
Various Deep Learning Approaches Have Been Applied Blockchain Intrusion detection
| Criteria | Blockchain Focus | Blockchain Integration | Main Technique | Dataset | Performance | Key Contributions | Weaknesses/Challenges |
|---|---|---|---|---|---|---|---|
| Al-Kadi et al. (2020) [50] | Ethereum | off-chain | BiLSTM for detecting intrusions | UNSW-NB15 and BoT-IoT | Most attack types detected with over 95% accuracy | Integrates deep learning and blockchain for secure VM migration and collaborative IDS | Potential scalability issues and complexity in deployment |
| Wang (2022) [52] | Embedded within architecture protocol | off-chain | Deep Residual Learning with edge info fusion | Not clearly specified | N/R | Uses deep residual fusion and association rules for embedded systems | Lack of standard datasets, unclear integration with broader blockchain systems |
| Mansour (2022) [54] | Blockchain for CPS communication and integrity | off-chain | AHSA for feature selection, ABi-GRNN, PRO optimizer | NSL-KDD-2015 CICIDS-2017 | N/R | Uses attention-based BiGRNN and hybrid optimization in a CPS blockchain context | generalization and reproducibility concerns |
| Katib & Ragab (2023) [53] | Generic blockchain framework | off-chain | HHO + SCA for feature selection, LSTM-AE, AOA for hyperparameter tuning | BoT-IoT | Accuracy: 99.05% | Hybrid metaheuristic optimization for feature selection; LSTM-AE for classification; Blockchain for secure IoT data flow | Model complexity and computation cost due to multiple optimization layers |
| Alamro et al. (2023) [56] | Used for secure medical data transmission | off-chain | ALO for FS, CNN + LSTM, FPA for tuning | ToN-IoT CICIDS-2017 | N/R | Applies hybrid optimization and DL in healthcare; leverages CNN-LSTM synergy | lacks interpretability and energy efficiency assessment |
| Kamali Poorazad et al. (2023) [57] | Generic Blockchain | off-chain | CNN-based IDS + Blockchain integration in SDN | Custom SDN attack simulation dataset (version 3) | N/R | Dual-layer security (IDS + blockchain); focuses on SDN-IIoT integration | Does not benchmark blockchain overhead or real-time latency |
| Li, Sun & Sun (2023) [58] | Used for privacy-preserving federated learning | off-chain | Federated LSTM-GRU with blockchain-secured training | NSL-KDD 2015 | Accuracy:99.01% | Focuses on privacy-preserving training; hybrid RNN + blockchain | Model complexity and hardware heterogeneity limit scalability |
| Alasmari (2024) [51] | Ethereum | off-chain | Ensemble of CNN, RBM, GAN with HSHBA optimization | UNSW-NB15 | Accuracy: 99.12%, Precision: 99% | Hybrid optimization (seahorse + bat), advanced ensemble DL models for IDS | High computational cost due to complex ensemble and optimization techniques |
| Alkhonaini et al. (2024) [55] | Blockchain for data sharing integrity | off-chain | Sandpiper optimizer, CNN-SAE, BSOA tuning | ToN-IoT, CICIDS-2017 | Accuracy: 99.59% (ToN-IoT), 99.54% (CICIDS-2017) | Efficient FS and hybrid DL for high accuracy; robust feature learning and hyperparameter tuning | Requires multiple tuning layers, increasing computational cost |
| Alruwaili (2025) [59] | blockchain-assisted IoT environment | Hybrid | GRU classifier; GJO feature selection; PDO-DE hyper-param tuning (LBCCD-GJO) | ToN-IoT | Up to 99.67% Acc | Robust DDoS pipeline with evolutionary optimization; compares processing time vs baselines | On-chain design details light; need device-level constraints and reproducibility checks. |
| Aliyu et al. (2025) [60] | Ethereum | Hybrid | LSTM core; self-updating NNs per block; | Binance Smart Chain, Ethereum Classic | Acc. 98.5%, FPR 1.5% | Integrate TEEs with blockchain for tamper-resistant, self-adapting IDS; | Trust/attacker model for TEEs and on-chain cost not fully benchmarked; |
Table 6.
Various Deep Learning Approaches Have Been Applied Blockchain Consensus Mechanisms
| Criteria | Blockchain Focus | Blockchain Integration | Main Technique | Dataset | Performance | Key Contributions | Weaknesses/Challenges |
|---|---|---|---|---|---|---|---|
| Liu et al. (2021) [65] | PoLe (Proof of Learning) | Hybrid | Neural network training with Secure Mapping Layer (SML) to prevent cheating | MNIST, IRIS, and CIFAR-10; data released to all network nodes | Stable block generation and efficient transaction processing shown in experiments | Prevents cheating with SML; redirects PoW effort toward useful neural network training | SML adds extra computational overhead; assumes honest behavior for validation |
| Badruddoja et al. (2022) [69] | Ethereum-like Smart Contract Blockchain | On-chain | On-chain DL Prediction via Taylor Series | MNIST | 99% accuracy; scalable on-chain prediction | Taylor-based approximation of activation functions for on-chain DL | Computational cost of on-chain operations; limited to simple models |
| Kim (2022) [68] | Custom LSTM Blockchain Framework | Hybrid | LSTM + Blockchain Consensus | Vehicle sensor and CAN data (simulated) | Real-time anomaly detection with secure propagation | Integrated LSTM for auto anomaly detection in vehicles | Specific to automotive use; lacks generalizability |
| Goh et al. (2022) [67] | Delegated PoS (e.g., EOS-like) | Hybrid | Deep Reinforcement Learning (D3P with TD-PBFT) | Simulation with varying node trust levels | High TPS, reduced computing cost, improved security | Trust-based node selection; DRL optimization; mitigates malicious influence | Focus on simulation; lacks real-world scalability validation |
| Islam et al. (2023) [66] | Generic PoS Blockchain | Hybrid | Multi-Agent Reinforcement Learning (MRL-PoS) | Simulated Blockchain Environment | Adaptive consensus; efficient malicious node detection | Introduced MRL-PoS; dynamic penalty-reward; RL-based voting system | No real-world deployment; complexity in multi-agent coordination |
| He et al. (2024) [64] | PPPML (Proof of Privacy-Preserving Machine Learning) | Hybrid | Privacy-preserving encrypted model training and hybrid learning | Confidential datasets; encryption techniques used to protect data privacy | Maintains model accuracy while preserving privacy; simulation validated | Introduces privacy protection for DL in blockchain consensus; hybrid approach balances privacy and efficiency | High computational cost of encrypted training; complex implementation |
| Alruwaili (2024) [70] | Custom IoT-Cloud Blockchain | Hybrid | FGADL-DEVCA (DL + Blockchain + Edge verification) | IoT-cloud simulation with injected faults | Enhanced fault detection; efficient edge validation | DAE for fault detection; Football Game Algorithm for tuning | Novelty in tuning may lack optimization efficiency; experimental |
| Zhi et al. (2025) [63] | BCDDL (Proof of Distributed Deep Learning Work) | Hybrid | Distributed Deep Learning used as mining task; dynamic incentives (DIM-TSCR) and model aggregation (MAA-TM) | Training data dynamically allocated based on node capability | Increases resource utilization and reduces impact of malicious nodes | Efficient use of resources in blockchain by integrating useful DL tasks; novel incentive and aggregation mechanisms | Requires reliable node performance estimation; potential vulnerabilities in data allocation |
| Amarnadh & Rao (2025) [71] | Byzantine lightweight fault-tolerant consensus | Hybrid | MRBSAN classifier; feature selection via enhanced Mountain Gazelle | Three credit datasets | Acc 97.2%, Prec 98.6%, Rec 98.6%, F1 98.0% | New MRBSAN architecture for credit-risk scoring | interpretability/XAI not discussed |
Table 7.
Various Deep Learning Approaches Have Been Applied Blockchain Data Privacy
| Criteria | Blockchain Focus | Blockchain Integration | Main Technique | Dataset | Performance | Key Contributions | Weaknesses/Challenges |
|---|---|---|---|---|---|---|---|
| Fadaeddini et al. (2019) [74] | Stellar | Hybrid | Decentralized federated learning using Stellar with Deep Learning Coin (DLC) incentive mechanism | General AI/Big Data (not specified) | Focus on incentivization and decentralization, no quantitative benchmarks provided | Introduced Stellar blockchain for deep learning model training with a novel token-based incentive mechanism (DLC) | Data availability, trust in validators, scalability of Stellar network |
| Zhu and Li (2021) [76] | Ethereum | Hybrid | Privacy-preserving federated deep learning with secure multi-party computation and Paillier encryption | not specified | Evaluated with respect to ciphertext size, throughput, training accuracy, and time | Proposed PDFDL model with encrypted gradient aggregation and smart contract-based model updates | Complex encryption adds overhead; does not fully guarantee privacy due to blockchain exposure risks |
| Zhou et al. (2022) [83] | Ethereum | Hybrid | GAN + Transfer Learning + Blockchain | EEG data (wearable sensors) | Secure, privacy-aware data sharing | BCI privacy preservation with synthetic data and blockchain | Limited application to small, specific use cases |
| Mahmood and Jusas (2022) [75] | Ethereum | Hybrid | Federated learning with multi-layered blockchain security and Zero-Knowledge Proofs (ZKPs) | Fashion-MNIST | Shows training accuracy and loss on CNN with Fashion-MNIST dataset | Integrated FL with blockchain to defend against multiple attacks and introduced multi-layered defense with ZKPs | FL still susceptible to attacks like poisoning and inference; complexity of securing large decentralized networks |
| Malik et al. (2023) [80] | Optimized Blockchain (Bonobo algorithm) | Hybrid | Feistel Encryption + DRL + Smart Governance | BoT-IoT, ToN-IoT | Enhanced security and fraud detection | Secure data exchange and governance using DRL and blockchain | Privacy vs. transparency trade-off |
| Awotunde et al. (2023) [77] | General Blockchain (ePoW) | Hybrid | Blockchain + CNN + KPCA | ToN-IoT, BoT-IoT | Improved accuracy | Hybrid deep learning and blockchain model for smart city security | High computational cost and complexity in multi-layer integration |
| Raju et al. (2023)[78] | Custom Blockchain | Hybrid | OK-HECCFHE + GRU + HPB- ASMO | IoT-Healthcare (custom data) | Secure prediction with encrypted cloud communication | Hybrid encryption and deep learning for healthcare IoT | Overhead from complex hybrid cryptographic scheme |
| Kokila & Reddy (2024) [79] | Ethereum | Hybrid | BlockDLO + CNN + Smart Contracts | Custom in-house testbed | High accuracy, reduced latency, lower energy use | Five-layer blockchain-deep learning model for IoT edge security | Complex architecture, limited generalizability |
| Li et al. (2024) [81] | Generic Blockchain | Hybrid | Probabilistic Encryption + ZKPs + GNN | Simulated transaction data | Balanced privacy and regulatory compliance | Framework for privacy-preserving, auditable blockchain transactions | Implementation complexity, not consensus-specific |
| Ghani et al. (2024) [82] | Decentralized Blockchain | Hybrid | GAN + Deep Learning + Blockchain | CelebA FFHQ HF | Improved privacy in biometric systems | Privacy-enhanced facial recognition on blockchain | Scalability and real-time performance concerns |
| Bhardwaj and Sumangali (2025) [85] | Generic Blockchain | Hybrid | Entropy Deep Belief Network (EDBN) within FL | Breast Cancer Wisconsin; UCI Heart | ~95–96% (BCW) and ~93% (Heart); | First end-to-end FL+ Blockchain+ XAI+ Optimization pipeline with auditable explanations | Added complexity; permissioned ledger + XAI + optimizer overhead; |
| Bezanjani et al. (2025) [84] | Generic Blockchain | Hybrid | LSTM for anomaly detection. | Simulation-based evaluation | +2% precision, accuracy, recall | proves feasibility of lightweight DL | Simulation; PoW— even lightweight—may add energy/latency overhead |
Table 8.
Various Deep Learning Approaches Have Been Applied Blockchain Malware Detection
| Criteria | Blockchain Focus | Blockchain Integration | Main Technique | Dataset | Performance | Key Contributions | Weaknesses/Challenges |
|---|---|---|---|---|---|---|---|
| Pastor et al. (2020) [91] | None (focuses on ML for encrypted traffic) | Off-chain | Random Forest, DNN with Tstat-derived flow features | Custom encrypted Monero mining traffic | N/R | Detection of encrypted cryptomining flows via network behavior profiling | No use of blockchain; dependent on realistic setup |
| Alotaibi (2021) [88] | Miyaguchi–Preneel cryptographic hash-based blockchain | Hybrid | Deep multilayer perceptive learning with Ruzicka index and biserial correlation | Drebin dataset | Accuracy 93–95% depending on sample size | Combines biserial correlation, blockchain hashing, and Ruzicka index for secure and fast malware detection in IoMT | Potential challenges in scalability for very large, real-time datasets were not discussedcomparison benchmarks |
| Nalinipriya et al. (2021) [89] | Blockchain network for decentralized security layer | Hybrid | Deep Stacked Autoencoder with Water Wave–based Moth Flame Optimization (WMFO) | Custom dataset using 2-gram and 3-gram opcodes | Accuracy: 96.93%, Sensitivity: 96.90%, Specificity: 97.92% | Uses optimized feature fusion and transformation methods for effective ransomware detection | Optimization complexity; performance might vary on real-world datasets |
| Denysiuk et al. (2023) | Blockchain with Proof-of-Action consensus | Hybrid | Deep learning + Blockchain consensus validation | Not specified | Accuracy: 98.81% to 99.33% | Parallel malware analysis via blockchain; Proof-of-Action validation | High computational complexity due to consensus process |
| Pawar et al. (2024) [90] | Blockchain for smart city authentication & data exchange | Hybrid | Two-layer XGBoost model with Bonobo optimization and greedy search | Raspberry Pi, NVIDIA Jetson, plus other IoT devices | F1-score: 93%, Accuracy: 95% | Integrates optimized XGBoost and blockchain to support digital governance security | Increased power and memory usage on edge devices (13.5% more energy) |
| Şafak et al. (2025) [92] | Blockchain (unspecified, used to store predictions immutably) | Hybrid | CNN-based ensemble (EfficientNetB0, MobileNetV2, Custom CNN) | CICMalDroid 2020 (3,590 image samples) | Accuracy: 97.38% | Image-based Android malware detection + blockchain logging | Does not specify blockchain type; mobile feasibility implied but not deeply tested |
| Nannan et al. (2025) [94] | Coordination and integrity of FL rounds | Hybrid | Federated Learning across devices | Not specified | N/R | Get hitched FL with blockchain to eliminate raw-data centralization | ledger throughput/latency and system complexity |