Monish
Saravanan

Traditional steganography methods like LSB insertion are easy to attack — they have low payload capacity and are detected quickly by steganalysis tools. This paper proposes a CNN-based encoder-decoder that hides data inside images with nearly undetectable visual change, trained on 10,000 images to handle real-world distortions like compression, noise, and scaling.

The encoder extracts deep spatial features from the cover image and embeds the secret payload across them. The decoder reverses this precisely, even after the image has been manipulated. A composite loss function balances two objectives simultaneously — keeping the image visually identical to the original (MSE) while maximising data recovery accuracy (Binary Cross-Entropy).

Steganalysis detection rate kept below 30% — meaning adversaries struggle to even confirm hidden data exists. The system also maintained 88% accuracy under Gaussian noise and 85% after 50% image scaling, with a maximum payload of 1KB per image.

Micro-plastics — particles smaller than 5mm — blend into sediment, surface reflections, and organic matter. They're translucent, tiny, and nearly impossible for a standard detection model to pick out in drone footage of a river.

I extended YOLOv8 with Convolutional Block Attention Modules (CBAM) inserted directly after the C2f backbone blocks, before the multi-scale fusion stage. This forces the network to suppress background noise — glare, ripples, colour variation — and amplify the faint signals left by translucent particles, rather than treating everything equally. Prior work only adds attention at the detection head; putting it earlier means the benefit carries through every downstream stage.

Training handled a sharp imbalance — micro-plastics are only 4.9% of annotations while macro-plastics account for 56%. I used focal loss to stop the model from ignoring them entirely, then oversampled micro-plastic examples with domain-specific augmentations: simulated murky water (brownish-green tint), Gaussian blur for camera defocus, and gamma correction to replicate changing light. YOLOv8's mosaic augmentation added further context diversity.

Defects can appear as changes in surface colour, texture, or shape. I'm investigating whether combining RGB images with 3D geometry improves industrial anomaly detection when only a limited number of defect-free training examples are available.

The study uses MVTec 3D-AD, initially focusing on bagel, carrot, and cookie objects. The planned approach aligns EfficientNet-B0 visual features with Fast Point Feature Histograms (FPFH) describing local 3D geometry. A PatchCore-style memory bank of normal features provides a reference for scoring new objects and locating potential defects.

Ongoing work. The evaluation will compare feature fusion with RGB-only, 3D-only, late-fusion, and RGB autoencoder baselines across three random seeds. Planned measures include image-level AUROC, pixel-level AUROC and AUPRO, runtime, and memory use. Results will be added after the experiments are complete.

A drop in detection precision does not always mean a model has stopped learning useful patterns. When illicit transactions become rarer, precision can fall even without a comparable change in the model's ability to distinguish them. I'm studying how to separate model degradation from changes in the underlying illicit-transaction rate.

Using the Elliptic transaction graph, the study compares Logistic Regression, Random Forest, a multilayer perceptron, and GraphSAGE to examine whether graph connections add value over transaction features. Chronological training, validation, and test windows keep future observations out of model fitting and tuning.

Ongoing work. The evaluation plan uses time steps 1–30 for training, 31–34 for validation, and 35–49 for testing. Per-step precision–recall performance and illicit-class F1 will be analysed alongside class prevalence, calibration, and bootstrap uncertainty. An unlabelled drift monitor is a stretch goal; final findings are not yet reported here.