Table 1 Summary of Recent Studies on Adversarial Attacks and Defense Techniques in Biometric, Medical, and Autonomous Systems.

⁸

EMG-based biometric ID

Adversarial style transfer, gradient attacks, universal perturbations, adversarial patches

Synthetic EMG signals deceive biometric models, exposing vulnerability

Biometric security (EMG, ECG, EEG)

Calls for improved security measures

⁹

EEG-based BCI

NPP backdoor attack

Embeds backdoor key in EEG signals, compromising BCIs

BCI

Need for enhanced security strategies

¹⁰

EEG-based BCI

ABAT (EEG data alignment)

Improves accuracy and robustness against adversarial attacks

BCI, autonomous driving, facial recognition

Data alignment for defense

¹¹

Adversarial attacks on ACO

Bias-based framework for adversarial patches

Exposes vulnerabilities in Ant Colony Optimization systems

Autonomous driving, facial recognition

Perceptual/attentional bias-based defense

¹²

Autonomous driving

Statistical mechanics-based model

Provides insights and mitigation strategies for adversarial robustness

Autonomous driving

Statistical mechanics model

¹³

kNN-based deep learning

ASK framework (differentiable loss)

Enhances attack/defense success rates in kNN-based models

Autonomous driving, facial recognition

Differentiable loss-based attacks/defense

¹⁴

kNN-based DNN models

ASK framework

Deepens understanding of adversarial strategies

Autonomous driving, facial recognition

Robust defense mechanisms

¹⁵

Bio signal classification

Early exit ensemble technique

Runtime robustness with less computational cost

Health-related models

Ensemble technique for runtime defense

¹⁶

Medical image analysis

FSRAA (Feature Space-Restricted Attention Attack)

Generates lesion-specific adversarial examples

Medical imaging, autonomous driving, facial recognition

Adversarial attack method

¹⁷

DNN adversarial defense

RAILS (immune-inspired evolutionary optimization)

Effective detection and defense against adversarial samples

Autonomous driving, facial recognition

Immune-inspired defense framework

¹⁸

Adversarial defense in DNNs

ANP (adversarial noise propagation)

Improves robustness with layer-wise noise injection

Autonomous driving, facial recognition

Noise injection in hidden layers

¹⁹

EEG-based BCI

UFGSM (unsupervised adversarial examples)

Effective adversarial examples generation without labels

EEG-based BCI

Calls for robust defenses

²⁰

DNN adversarial robustness

SNS (sensitivity-based neuron selection)

Enhances robustness via neuron sensitivity

Autonomous driving, facial recognition

Neuron sensitivity-based defense

²¹

Medical image (ultrasound)

Novel adversarial attack on image reconstruction

48% misclassification in fatty liver disease diagnosis

Medical imaging

Emphasizes robust training data

²²

COVID-19 detection

Property inference attacks

Highlights risks of information leakage from ML models

Healthcare

Privacy-preserving techniques

²³

Medical imaging adversarial attacks

Hierarchical feature constraint (HFC)

Hides adversarial features in target distribution, evades detectors

Clinical decision-making

Improved detector robustness

²⁴

Medical diagnostic model defense

MedRDF (majority voting on noisy copies)

Enhances robustness at inference without retraining

COVID-19, DermaMNIST

No retraining, inference-time defense

²⁵

EEG seizure detection

Gaussian-Stockwell Transform + Hermite Polynomial Features

Improved accuracy, sensitivity, and specificity for seizures

Neurological disorder diagnosis

Advanced ML feature engineering

²⁶

EEG motor imagery classification

Attention-based deep learning model

High-performance but vulnerable to adversarial attacks

EEG-based motor imagery

Need for stronger robustness

²⁷

EEG seizure detection

CNN + Explainable AI (SHAP, LIME)

High accuracy (98.08%) with interpretability

Epilepsy diagnosis

Explainable AI enhances trust