THE RESEARCH LAB / M.S. COMPUTER ENGINEERING
Seeing beyond
the visible.
An image can look unchanged and still carry a different signal. Stego-BT asks whether a model can learn to ignore that hidden disturbance.
INTERACTIVE / THE SMALLEST POSSIBLE CHANGE
Flip a bit.
See the difference.
Each 8-bit pixel channel is a number from 0 to 255. Changing its least significant bit changes that number by just one. A model still receives a different input.
Educational simulation on a synthetic pixel pattern, not an experiment result. The research’s LSB case is plane 0; higher planes illustrate why bit position matters.
Each layer contributes a different weight to the same pixel.
Clean + modified views → shared encoder ↓01011010→0101101190 → 91 · difference of 1
01 / THE VULNERABILITY
Robust to crops.
But to hidden bits?
Self-supervised learning teaches an encoder to recognize that different views come from the same image. Familiar augmentations such as crops and color changes encourage useful invariances.
My capstone investigates a different disturbance: steganographic changes to low-order pixel bits. Stego-BT brings clean and steganographically modified views into training so the learned representation can become less sensitive to that signal.
02 / THE TRAINING PIPELINE
Different views. Shared meaning.
The reported run used an 800-epoch CIFAR-10 schedule with mixed precision on NVIDIA T4 hardware, reaching approximately 1.8× throughput within a 16GB memory budget.
- 01 / INPUT
CIFAR-10 image
Start with the same underlying example.
- 02 / VIEWS
Clean + stego
Heavy augmentation and LSB bit-plane modification.
- 03 / ENCODER
Shared representation
Learn image features across the views.
- 04 / OBJECTIVE
Barlow Twins
Align corresponding features; reduce redundancy.
Why redundancy reduction?
The cross-correlation matrix should approach identity: matching feature dimensions agree across views, while different dimensions avoid repeating the same information.
L = Σᵢ (1 − Cᵢᵢ)² + λ Σᵢ Σⱼ≠ᵢ Cᵢⱼ²C is cross-correlation between representations. The first term encourages agreement; the second penalizes redundant dimensions.
03 / EXPERIMENT DESIGN
Two-stage fine-tuning
First, freeze the encoder and fit a linear probe to evaluate its features. Then fine-tune the full backbone with discriminative learning rates of 10⁻⁴ and 10⁻⁵.
I trained for 800 epochs with mixed precision on NVIDIA T4 GPUs.
Four-view ablation
Add local crops alongside clean and stego global views to investigate how spatial granularity changes the representation.
The question: does learning from local detail change robustness to a global hidden signal?
04 / REPORTED RESULTS
Resilience worth investigating.
Robustness under LSB attack
Stego-BT reduced the reported degradation under attack to under 1%, compared with a six-point drop in the stated baseline. Learning from clean and stego views made resilience part of the objective.
Compute comparison
1.5G FLOPs versus 4.1G for the ResNet-18 baseline in the original evaluation setup.
Research evidence & provenance
Results and methods are recorded in the original project materials summarized in my resumes. The training run was not reproduced during preparation of this portfolio; the thesis and experiment artifacts were unavailable. No full attack-strength curve or ablation result is inferred. The interactive illustration uses synthetic pixels and is not a benchmark.