Advanced Research Topics (7PAM2016) — Building GANs from Scratch and Applying Them to Medical Imaging, Network Traffic and Sketch Generation
This Masters-level assessment asks for a complete generative adversarial network study, delivered as an annotated code submission carrying sixty per cent of the marks and a six-to-eight page technical report carrying the remaining forty. The work spans four separate GAN implementations, moving from a controlled synthetic setting into three contrasting real-world application domains. Part one builds a GAN from scratch in PyTorch on synthetic two-dimensional data. The tutorial sine-wave generator is reproduced first as a baseline, then a new distribution is modelled — a noisy parametric curve of the form y = sin(2x) + 0.3cos(5x) with an additive noise term — before the architecture itself is varied. Activation functions and layer depth are altered systematically and the resulting sample distributions plotted against the originals, so the effect of each architectural choice on convergence and sample fidelity can be seen rather than asserted. Part two applies the same principles at scale across three domains. The medical strand trains a DCGAN on the OCTMNIST subset of MedMNIST, generating synthetic optical coherence tomography retinal images, tracking generator and discriminator losses across training, and evaluating output both visually and quantitatively using Fréchet Inception Distance. A conditional GAN extension conditions the generator on class label so that images for a chosen retinal pathology can be produced on demand. The cybersecurity strand shifts from images to feature vectors, using preprocessed CICIDS 2017 network intrusion data. Benign and DoS traffic is combined and explored for class balance, a GAN is built to synthesise tabular feature vectors rather than pixels, and real against generated distributions are compared through PCA and t-SNE projections, with a discussion of how well the model generalises across attack types. The creative strand trains a DCGAN on the QuickDraw 'birthday cake' sketch category, tracking visual outputs epoch by epoch and benchmarking generated sketches against real ones, with an extension covering additional categories of differing sketch complexity. The accompanying report explains the analysis steps and the reasoning behind each architectural decision rather than restating textbook definitions of the method. It gives brief descriptions of the models used, presents generated samples and loss curves as figures, interprets the evaluation metrics, and reflects honestly on failure modes — training instability, mode collapse, and the visible flaws in synthetic output that determine whether such data is fit for downstream use. The code is written as reusable functions, commented for a reader other than its author, and reproduces every figure and numerical value quoted in the report.
Read Model Answer →