Kawser Sikder
Contact
Contact

Research

Trustworthy evaluation in deepfake forensics and computational pathology

Two projects, one question: does a reported result still hold under stricter evaluation? Both use pre-registered analyses and confidence intervals, report negative results as they came out, and have public code.

PythonPyTorchscikit-learnVision transformers · LoRA · Survival analysis

  • Undergraduate thesis

    CSE 465

    Cross-dataset deepfake detection with foundation models

    North South University, June 2026 to present. Two-person senior design project with T. A. Aurnobb, supervised by Dr. Riasat Khan. Manuscript in preparation.

    Do deepfake detectors still work on datasets and compression levels they never saw in training? A controlled empirical study of foundation-model detectors under real-world distribution shift, guided by an 82-paper literature gap analysis.

    0.889
    Mean AUC on unseen datasets (EVA-02 + LoRA)
    +0.07 to +0.09
    AUC over a CNN baseline, every unseen set
    ~0.5%
    Parameters trained
    ~200
    GPU-hours on free Kaggle T4s
    Data
    FaceForensics++, Celeb-DF and DFDC. Train on two, test on the unseen third.
    Approach
    • LoRA-adapted CLIP ViT-L/14, DINOv2 ViT-g/14 and EVA-02-L/14, training about 0.5% of the parameters.
    • A controlled study of codec-aware priors for compression robustness, with matched-strength controls and real-codec baselines. Contrary to our hypothesis, codec-shaped masks did no better than random DCT masking (+0.045 AUC under heavy H.264, against +0.025 for training on real H.264; preliminary).
    • Pre-registered decision rules, paired cluster-bootstrap confidence intervals and measured run-to-run noise.
    • Found and fixed a silent reproducibility bug (training data depended on file listing order) by fingerprinting the data.
    • Fit a 1.1B-parameter DINOv2-g onto free 2 × T4 GPUs with frozen fp16 weights, about 200 GPU-hours on Kaggle.
    • Ongoing: frequency and attention analysis under compression, calibrated abstention, unseen codecs (HEVC, AV1) and new generators (SD3, FLUX.1).
    • Python
    • PyTorch
    • scikit-learn
    Code

  • Independent research

    Do H&E image features predict lymphoma survival beyond clinical scores?

    Self-directed, August 2026 to present. Computational pathology for diffuse large B-cell lymphoma (DLBCL). Manuscript drafted, not yet posted.

    A pre-registered test of whether nuclear morphology from routine H&E slides adds prognostic value beyond the IPI clinical score. It does not: the null result held across 3 cohorts, 4 feature views, 2 learners and 5 sensitivity analyses. A second study audits patient overlap between public DLBCL cohorts.

    1.96M
    Nuclei segmented from 170 patients
    0.888
    Segmentation Dice on LyNSeC
    −0.023
    C-index change vs. IPI (95% CI −0.050 to +0.001)
    39 of 66
    Papers validating on an overlapping cohort
    Data
    DLBCL-Morph (TCIA): 30,007 H&E patches from 170 patients. Segmentation checked on LyNSeC. Public GEO gene-expression cohorts for the overlap audit.
    Approach
    • U-Net + watershed nucleus segmentation, then nuclear pleomorphism, packing and patch-heterogeneity features for about 1.96 million nuclei.
    • Analysis protocol signed and hash-locked before any outcome data was read.
    • Penalised Cox and Random Survival Forest under 10 × 5 nested cross-validation, with paired bootstrap against the clinical model.
    • Reproducibility checks: a published DLBCL-Morph baseline fell from C = 0.635 to 0.521 under fold-safe refitting, a released segmentation checkpoint scored Dice 0.282 on its own source data, and my pipeline reproduced across two environments to within 1.8 × 10⁻¹⁴.
    • Ongoing cohort-overlap audit: up to 96% shared samples between widely used GEO cohorts. Of 66 open-access papers using an overlapping pair, 39 trained on one and validated on the other.
    • Python
    • PyTorch
    • scikit-learn
    Code

Foundations

The coursework behind it

From supervised and unsupervised learning to CNNs, RNNs and Transformers, with the official outline and what my section covered for each.

88%
CSE 445

Machine Learning

Supervised learning algorithms across the board, plus core unsupervised methods.

B+ · 3.3Pythonscikit-learn