SkinGuardian: On-Device AI for Private, Fair, Robust, and Explainable Skin Cancer Detection

Background: Early skin cancer detection improves outcomes, but access to dermatology screening remains limited. Many AI systems rely on cloud inference, raising privacy concerns and restricting use in low-connectivity settings. Methods: We present SkinGuardian, an on-device benign–malignant skin lesion classifier that integrates four trustworthiness dimensions: fairnessaware learning, adversarial robustness, differential privacy, and explainability. We fine-tune a BEiT vision transformer on ISIC 2019 and Fitzpatrick17k (train/validation only; test held out for subgroup evaluation), and deploy via ONNX Runtime with INT8 weights-only quantization. Results: SkinGuardian-Clean achieves AUROC 0.956 on ISIC 2019, and generalizes to the SIIM-ISIC 2020 melanoma setting with AUROC 0.927; at the ISIC-2019-tuned operating threshold, accuracy is 85.4%. Fairness mitigation reduces demographic parity difference on Fitzpatrick17k from 0.12 to 0.04 and equalized odds difference from 0.15 to 0.05. SkinGuardian-Robust attains 74.8% robust accuracy against PGD-10 (ϵ = 8/255; clean 87.1%). With DPSGD, accuracy remains 86.1% at ϵ = 1 (δ = 1/N) on ISIC 2019. On-device inference achieves p95 ≤160 ms with INT8. Conclusion: SkinGuardian demonstrates a practical, privacy-preserving and equitable on-device screening research prototype and is not a standalone diagnostic device.