Digital Signal Processing and Artificial Intelligence for Automatic Learning K Kiruthika 5 3 2026 https://doi.org/10.6025/dspaial/2026/5/3/168-186 https://www.dline.info/dspai/fulltext/v5n3/dspaiv5n3_2.pdf Handwritten digit recognition systems are often expected to perform reliably under imperfect input conditions, yet simple neural architectures are frequently evaluated only on clean benchmark data. This study systematically evaluates the robustness of a multilayer perceptron (MLP) trained on the MNIST handwritten digit dataset under four controlled perturbation families: additive Gaussian noise, salt and pepper impulse noise, contiguous pixel masking, and spatial translation. The trained MLP achieves a clean test accuracy of 92.75%. Perturbation analysis using accuracy, precision, recall, F1-score, confusion matrices, and accuracyversus- strength curves shows that the model is comparatively resilient to sparse and localized degradations, maintaining 86.40% accuracy under salt and pepper noise and 82.95% under pixel masking. Performance decreases more noticeably under Gaussian noise, with accuracy falling to 77.35%. In contrast, even small translations cause severe degradation, with accuracy dropping to 33.35% at a 2-pixel shift and below 5% at shifts of 4–5 pixels. Gradient-based saliency maps reveal that the network relies on localized stroke pixels at fixed spatial positions, explaining its pronounced vulnerability to translation. The findings demonstrate that MLP robustness is highly perturbation-specific and fundamentally constrained by the absence of spatial inductive biases such as weight sharing and local connectivity. The study provides practical guidance for deploying lightweight classifiers and highlights the importance of preprocessing or translation invariant architectures in real world handwriting recognition applications.