MECP-GAP: Mobility-Aware MEC Planning via GNNs
Auditing Backdoors: From a Flag to a Causal Proof
Overview
A model that gets flagged as “probably backdoored” is not the same as a model whose exact trigger and payload have been proven, causally, to be connected. This project builds on a published pipeline for extracting and reconstructing hidden LLM backdoor triggers, and adds a downstream layer whose whole purpose is to close the gap between “looks suspicious” and “verified.”
Method
The base pipeline probes a suspect model without any prior knowledge of its trigger, clusters what leaks out into recurring patterns, and searches for the exact trigger and target from those patterns. On top of that, a second auditing stage treats a candidate trigger as something to be tested causally against a matched neutral control, rather than accepted on similarity alone. A further layer sits above the auditor and decides, under a limited budget of queries to the suspect model, which experiment to run next.
Results
Running the base pipeline at scale surfaced an important failure mode: it flagged the large majority of backdoored models correctly, but the string it actually recovered as the “target” was frequently generic boilerplate text rather than the real payload — a reminder that a high detection rate and a correct explanation are two different claims. The stricter causal layer was built specifically to catch that gap. A budget-aware policy for choosing what to test next cut the number of queries needed to confirm a real trigger–target link by more than half compared to testing everything uniformly — though a deeper follow-up study found that the reason it worked was not the one first assumed, which is now the main open question before the method is run against real, uncontrolled models.
2026
2025
A Watchdog Model for Backdoored LoRA Adapters
Overview Most simple backdoor checks look for a specific trigger phrase. That breaks the moment the trigger is unknown or paraphrased. This project asks a di...
The Sentinel’s Dilemma: Guarding AI from Hidden Threats
The Sentinel’s Dilemma: Who Guards Our AI Guardians?
Turning a Backdoor’s Step Function Into a Ramp
Overview A backdoor behaves like a step function: invisible until its trigger fires, fully present the instant it does. That makes it a bad target for anythi...
Auditing Backdoors: From a Flag to a Causal Proof
Overview A model that gets flagged as “probably backdoored” is not the same as a model whose exact trigger and payload have been proven, causally, to be conn...
Re-running a Vulnerability Detection Benchmark, More Carefully
Overview A widely cited graph neural network for detecting vulnerable code is reproduced from scratch on the authors’ own released data, then put under three...
Toward a Field-Based View of LLM Backdoors
Overview Existing LLM backdoor scanners mostly search for an object: a trigger string, a target string, or a memorized leaked example. This project explores ...
Building a Weakness Zoo to Stress-Test a Backdoor Scanner
Overview A backdoor scanner that only gets tested against easy, textbook backdoors will look better than it is. This project builds a “weakness zoo” — a deli...
Starting Point: Vision-Language Models for Dental X-rays
Overview Dental panoramic X-rays are read manually by clinicians, and there is growing interest in whether general-purpose medical vision-language models can...
Comparing Five Ways to Catch a Backdoor, Live
Overview Most backdoor-detection papers evaluate on large models behind a GPU cluster, which makes it hard to build intuition about when a detection signal a...
Understanding BackdoorBench: A Comprehensive Benchmark for AI Security
Unveiling BackdoorBench: A Critical Benchmark for AI Security
2024
Wav2Vec2 Sentiment Analysis Using Shemo Dataset
Overview In this project, we fine-tuned the Wav2Vec2 model to perform sentiment analysis based on both voice features and text transcripts from the Shemo da...
Osmium Project Tasks
Overview In this project, we developed an Android application to estimate the location of cellular network cells using Received Signal Strength Indicator (RS...
Multimodal Sentiment Analysis Project(Persian-3classes)
Overview In this project, we explore multimodal sentiment analysis, which involves analyzing both text and image data together. Our goal is to predict sentim...
2023
Project Iridium
Project Iridium
My Internship at the NLP Lab
My Internship at the NLP Lab :
2021
My Heroku project with blazore
Its My Heroku project :
Sonic pi
My sonic pi project : . این پست من مربوط به پوروژه سونیک پای بنده است
2020
مصاحبه
My works and wishes Success secret مصاحبه با جناب اقا پارسا: ایشون بسیار ادم سخت کوشی بودن و گفتن که حتی در دوران دانشجویی شون درس هم میدادن 1 به موار...
My works and wishes
Its my favorite university This is oxford: In England The best unniversity. this university is first inthe world. this university have best descover...
My web site in lab exam(hacketan)
Its My Hackathon project :