Analisis Robustness Model Convolutional Neural Network terhadap Serangan Adversarial pada Skenario Man-In-The-Middle untuk Klasifikasi Pneumonia Berbasis Chest X-Ray

Kharisma Sekarningtyas, Ir. Yuris Mulya Saputra, S.T., M.Sc., Ph.D., IPM.

2026 | Tugas Akhir | D4 TEKNOLOGI JARINGAN

Pneumonia merupakan penyebab utama kematian akibat infeksi global, sehingga diagnosis akurat menjadi sangat krusial. Meskipun deep learning telah menunjukkan potensi besar sebagai sistem pendukung keputusan klinis, model ini rentan terhadap serangan adversarial. Penelitian ini bertujuan untuk menganalisis robustness dari model EfficientNetB0, DenseNet121, dan ResNet50 terhadap serangan FGSM dan PGD pada sistem klasifikasi pneumonia berbasis chest X-ray. Evaluasi dilakukan melalui offline evaluation dan skenario Man-in-the-Middle (MITM) pada arsitektur client-server dengan menggunakan metrik accuracy, precision, recall, F1-sore, latency, dan throughput. Hasil menunjukkan akurasi pada data independen berkisar 85%–87%, lebih rendah dibandingkan akurasi validasi (>98%), yang menegaskan pentingnya proses pembagian dataset yang tepat untuk memperoleh estimasi generalisasi. DenseNet121 terbukti memiliki tingkat robustness terbaik, sementara EfficientNetB0 paling rentan terhadap perturbasi. Serangan PGD menurunkan akurasi hingga 0% pada offline evaluation, sedangkan pada skenario MITM akurasi masih bertahan pada kisaran 40%–65% akibat transferability gap. Skenario MITM juga meningkatkan latency lebih dari 10 kali lipat serta menurunkan throughput sebesar 90%–94% akibat proses pembentukan perturbasi. Temuan ini menunjukkan bahwa evaluasi akurasi saja belum memadai sehingga evaluasi robustness dan keamanan implementasi perlu dipertimbangkan sebelum sistem diterapkan pada lingkungan klinis.

Pneumonia is the leading cause of death from infections globally, making accurate diagnosis crucial. Although deep learning has shown great potential as a clinical decision support system, these models are vulnerable to adversarial attacks. This study aims to analyze the robustness of the EfficientNetB0, DenseNet121, and ResNet50 models against FGSM and PGD attacks on a chest X-ray-based pneumonia classification system. The evaluation was conducted thru offline evaluation and Man-in-the-Middle (MITM) scenarios on a client-server architecture using the metrics of accuracy, precision, recall, F1-score, latency, and throughput. The results show that the accuracy on independent data ranges from 85% to 87%, lower than the validation accuracy (>98%), highlighting the importance of proper dataset partitioning to obtain generalization estimates. DenseNet121 proved to have the best robustness level, while EfficientNetB0 was the most vulnerable to perturbations. The PGD attack reduces accuracy to 0% in the offline evaluation, while in the MITM scenario, accuracy remains in the range of 40%–65% due to the transferability gap. The MITM scenario also increases latency by more than 10 times and decreases throughput by 90%–94% due to the perturbation formation process. These findings indicate that accuracy evaluation alone is insufficient, and therefore, robustness and implementation security evaluations need to be considered before the system is applied in a clinical environment.

Kata Kunci : Pneumonia, Chest X-ray, Adversarial Attack, Man-in-the-Middle, Robustness.