Ali Hussain

B.S. Computer Science, NUST (2023 – 2027)

Islamabad, Pakistan

ahussain.bscs23seecs@seecs.edu.pk

I am a final-year Computer Science undergraduate at the National University of Sciences and Technology (NUST), Pakistan, and I am looking for a PhD position starting Fall 2027.

I’m drawn to a simple question: what are models actually relying on, and how do we find out? My favourite result so far is that document-forgery detectors were reading the JPEG compression grid rather than the forgery. Shifting the grid by 4 pixels, with no change to content or forgery, dropped a state-of-the-art detector from 0.73 to 0.11 F1. (one of three co-first authors, BMVC 2026 oral, to appear) I like experiments like that one: change something that shouldn’t matter, and see what breaks.

I’m also interested in making large models useful where data is scarce. For low-resource Urdu handwriting, I turned pen dynamics into image channels that a pretrained recognizer could use, reaching 3.66% CER against 7.28% for a QLoRA-fine-tuned Qwen3-VL-4B on the same test set. (first author, ICDAR 2026)

Lately I’ve been curious about generative and multimodal models: training them against verifiable rewards (my current work at DFKI) and what they lose when they are compressed (my final-year project with RPTU). (Directions I’m exploring)

I have worked with Prof. Faisal Shafait at NUST, with Prof. Andreas Dengel and Dr. Sheraz Ahmed at DFKI, and with Dr. Christian Weis at RPTU Kaiserslautern-Landau.

Things I’m curious about

Shortcuts and evaluation: what models really use, and benchmarks that can tell
Multimodal and generative models, especially for documents and text in images
Learning from verifiable rewards
Efficient models: distillation, compression and low-resource settings

I care about careful experimental practice: multiple seeds, significance tests, controlled ablations, and public code and data.

Outside research, I worked part-time as a computer vision engineer at CCRIPT, taking detection and tracking models from experiment to deployment on constrained hardware.

selected publications

ICDAR
Online Urdu Text-Line Recognition by Bridging Stroke Dynamics and Offline Representations

Ali Hussain, Rafay Ahmad, Momina Moetesum, Adnan Ul-Hasan, and Faisal Shafait

In Document Analysis and Recognition – ICDAR 2026, Proceedings, Part II. Dataset: OUHD-L on Zenodo , 2026

Abs DOI Bib PDF Code

Robust recognition of Urdu has been challenging due to its cursive structure and extensive ligature variations. Modern pen devices capture rich spatio-temporal signals; however, state-of-the-art Urdu research predominantly focuses on offline images, and online-first transformer and vision-language approaches overfit in low-resource settings. We present two contributions: (i) OUHD-L, a new dataset of 2,403 online Urdu handwritten text lines from 311 writers, captured with the Wacom IoT paper device at fine-grained pen-trajectory resolution; and (ii) an encoding of online information as image channels that lets us reuse pre-trained offline Urdu text-line recognizers. Fusing online-derived feature images reduces character error rate from 4.21% (ink-only) to 3.66%, showing that converting online dynamics into image space bridges the data gap while reusing mature offline architectures.
@inproceedings{hussain2026online, title = {Online Urdu Text-Line Recognition by Bridging Stroke Dynamics and Offline Representations}, author = {Hussain, Ali and Ahmad, Rafay and Moetesum, Momina and Ul-Hasan, Adnan and Shafait, Faisal}, booktitle = {Document Analysis and Recognition -- ICDAR 2026, Proceedings, Part II}, editor = {Fink, Gernot A. and others}, series = {Lecture Notes in Computer Science}, volume = {16973}, pages = {678--694}, publisher = {Springer Nature Switzerland}, address = {Cham}, year = {2026}, doi = {10.1007/978-3-032-36033-5_40}, url = {https://link.springer.com/chapter/10.1007/978-3-032-36033-5_40}, }
BMVC
Mistaking Periodicity for Manipulation: JPEG-Induced Structural Bias in Document Tampering Detection

Nauman Riaz^*, Ali Hussain^*, Saifullah Saifullah^*, Muhammad Imran Malik, Stefan Agne, Andreas Dengel, and Sheraz Ahmed

In British Machine Vision Conference (BMVC), 2026

Oral presentation. Accepted, to appear

Oral Abs Bib PDF

Selected for oral presentation at BMVC 2026.

Deep models for document tampering detection increasingly rely on multimodal RGB+DCT architectures, implicitly assuming that JPEG block artifact grids (BAGs) provide stable cross-forgery cues. We show that this assumption embeds a strong inductive bias that fails under minimal, adversarially constructed perturbations. Unlike natural images, document images contain sharply bounded glyph structures, making grid-aligned manipulations trivial for an adversary. We formalize this through two complementary attacks. Grid-Aligned Forgery (GAF) preserves local JPEG block statistics by aligning copy-move, splicing, or generative manipulations to the 8x8 grid. Pad-Recompress-Crop (PRC) globally shifts the JPEG grid while preserving semantic content and geometry. We introduce Detection Failure Rate (DFR) and False Positive Area (FPA) to capture failure modes not measured by prior work. On DocTamper and T-SROIE, both attacks substantially degrade state-of-the-art and robustness-oriented detectors (CAT-Net, DTD, FFDN, DocForgeNet, ADCD-Net), indicating a strong bias toward JPEG-grid statistics.
@inproceedings{riaz2026periodicity, title = {Mistaking Periodicity for Manipulation: JPEG-Induced Structural Bias in Document Tampering Detection}, author = {Riaz, Nauman and Hussain, Ali and Saifullah, Saifullah and Malik, Muhammad Imran and Agne, Stefan and Dengel, Andreas and Ahmed, Sheraz}, booktitle = {British Machine Vision Conference (BMVC)}, year = {2026}, note = {Oral presentation. Accepted, to appear}, }

news

Sep 05, 2026	Completed my research internship at RPTU Kaiserslautern-Landau (8 Jun – 5 Sep 2026). The distillation work continues as my final-year project.
Aug 31, 2026	My first-author paper on online Urdu text-line recognition was presented as a poster at ICDAR 2026 in Vienna and is published in the Springer LNCS proceedings. Code and the OUHD-L dataset are public.
Aug 07, 2026	Our paper Mistaking Periodicity for Manipulation was accepted as an oral at BMVC 2026. We show that document-tampering detectors rely on the JPEG 8×8 grid, and we introduce two attacks that break them.
Jun 08, 2026	Started a research internship at RPTU Kaiserslautern-Landau with Dr. Christian Weis on foundation-model distillation for edge inference.
Aug 31, 2025	Wrapped up my research internship at DFKI (1 Jun – 31 Aug 2025). This work led to our paper, accepted at BMVC 2026, on JPEG-induced bias in document tampering detection.
Jun 01, 2025	Started a research internship at DFKI in Kaiserslautern, Germany, on document image forensics with Prof. Andreas Dengel and Dr. Sheraz Ahmed.
Jun 01, 2024	Joined the TUKL-NUST R&D Center as an undergraduate researcher working on handwriting recognition with Prof. Faisal Shafait.