Impact of Adversarial Attacks on Robust Cancer Detection in Medical Images using Improved Optimization-based Adaptive Multi-Scale ShuffleNetV2

The uncontrolled proliferation of abnormal cells within any region of the body is called cancer. A collection of disorders brought on by aberrant cell growth in various sections of the body is together referred to as cancer. More than one hundred forms of cancer exist, including skin, lung, breast, oral, colon, and prostate cancer. Treatment delays might result in major health problems or even death. Because they are lethal, illnesses including cancer are commonly referred to as chronically fatal illnesses. Cancer is said to be deadly since it spreads more quickly and is typically discovered at an advanced stage. It has been discovered that reducing deaths from cancer requires early identification. However, new research has revealed that current models are susceptible to adversarial attack, which alters images by making tiny pixel-level adjustments to make algorithms incorrectly identify images. This paper presents Adaptive Multi-scale ShuffleNetV2 (AMSNetV2), a multi-scale adaptive evolution of ShuffleNetV2, designed to ensure reliable cancer detection even under adversarial manipulation. These attacks aim to deceive algorithms by making subtle alterations to images, potentially leading to misinterpretations in cancer detection. AMSNetV2 methodology is specifically tailored to combat such adversarial threats, ensuring the model’s reliability and accuracy in identifying cancerous regions. AMSNetV2’s strength lies in its parameter optimization using the Updated Random Value-based Zebra Optimization Algorithm (URV-ZOA). This optimization process fine-tunes the model’s parameters, enhancing its ability to discern cancer indicators within medical images even amidst adversarial manipulations. By leveraging URV-ZOA, AMSNetV2 can maintain its high-performance standards and robustness, crucial for accurate cancer detection. The multi-scale nature of AMSNetV2 further fortifies its defenses against adversarial attacks. This approach allows the network to detect fine-grained anomalies and complex textures that often signal the early stages of malignancy. By incorporating multi-scale analysis, AMSNetV2 enhances its resilience to adversarial perturbations, ensuring that its cancer detection capabilities remain steadfast and dependable. In the realm of medical imaging, where precision and reliability are paramount, AMSNetV2’s ability to withstand adversarial attacks while upholding detection accuracy is a significant advancement. Its adaptability and resistance to deceptive manipulations make it a valuable asset in the fight against cancer, providing medical professionals a high-integrity diagnostic tool for precise and consistent identification of cancer within medical imagery. The results proved that the designed AMSNetV2 attained the enhanced classification accuracies of 98.63% and 97.94% on the MIAS and CBIS-DDSM datasets. Moreover, under adversarial conditions, the proposed framework maintained robust accuracies of 96.84%, 94.57%, and 92.89% against Fast Gradient Sign Method (FGSM), Basic Iterative Method (BIM), and Projected Gradient Descent (PGD) attacks, respectively, on Dataset-1, while achieving 96.21%, 93.82%, and 91.94% on Dataset-2, thereby demonstrating its effectiveness and robustness for reliable cancer detection.