Research write-ups by Amirreza Vishteh on AI safety, LLM security, and NLP.https://www.amirrezavishteh.ir/https://www.amirrezavishteh.ir/blog/devign-reproduction/https://www.amirrezavishteh.ir/blog/devign-reproduction/I reproduced Devign (NeurIPS 2019) from scratch on the authors' own data. It lands 11 points below the paper, two-thirds of its test set leaks through shared commits, and a plain CNN beats it.Wed, 26 Aug 2026 00:00:00 GMTreproducibilitygraph neural networkssecurityhttps://www.amirrezavishteh.ir/blog/mecp-gap/https://www.amirrezavishteh.ir/blog/mecp-gap/A review and implementation of MECP-GAP, a graph-neural-network approach to partitioning 5G base stations into MEC service regions that minimizes handover cost under load-balancing constraints.Sat, 14 Feb 2026 00:00:00 GMTpaper reviewgraph neural networksnetworkshttps://www.amirrezavishteh.ir/blog/watchdog-for-backdoored-lora/https://www.amirrezavishteh.ir/blog/watchdog-for-backdoored-lora/Training a small distilled model to recognize the statistical fingerprint of a backdoored language model, instead of searching for a known trigger word.Sun, 16 Nov 2025 00:00:00 GMTAI safetybackdoorsLLMhttps://www.amirrezavishteh.ir/blog/the-sentinels-dilemma/https://www.amirrezavishteh.ir/blog/the-sentinels-dilemma/A survey of backdoor attacks on large language models — how hidden triggers are implanted during training, and the detection and defense strategies emerging against them.Tue, 14 Oct 2025 00:00:00 GMTAI safetybackdoorsLLMsurveyhttps://www.amirrezavishteh.ir/blog/backdoor-step-function-to-ramp/https://www.amirrezavishteh.ir/blog/backdoor-step-function-to-ramp/A research design for eliciting a hidden model behavior by amplifying the exact fine-tuning change that planted it, then dialing that amplification back down to isolate the trigger.Thu, 11 Sep 2025 00:00:00 GMTAI safetybackdoorsLLMhttps://www.amirrezavishteh.ir/blog/auditing-llm-backdoors/https://www.amirrezavishteh.ir/blog/auditing-llm-backdoors/Extending a trigger-reconstruction pipeline for fine-tuned LLMs with a stricter auditing layer that separates a plausible-looking flag from actual proof of a backdoor.Sun, 31 Aug 2025 00:00:00 GMTAI safetybackdoorsLLMhttps://www.amirrezavishteh.ir/blog/field-based-view-of-llm-backdoors/https://www.amirrezavishteh.ir/blog/field-based-view-of-llm-backdoors/A design study exploring whether a backdoor can be caught by looking at everything fine-tuning changed, rather than searching for a trigger or a target string.Tue, 19 Aug 2025 00:00:00 GMTAI safetybackdoorsLLMhttps://www.amirrezavishteh.ir/blog/bait-weakness-zoo/https://www.amirrezavishteh.ir/blog/bait-weakness-zoo/Deliberately building harder backdoor variants to find the blind spots of a state-of-the-art LLM backdoor scanner.Fri, 15 Aug 2025 00:00:00 GMTAI safetybackdoorsLLMhttps://www.amirrezavishteh.ir/blog/dental-xray-vision-language-models/https://www.amirrezavishteh.ir/blog/dental-xray-vision-language-models/Early exploration of applying medical vision-language models to dental panoramic radiographs.Tue, 24 Jun 2025 00:00:00 GMThealthcare AIcomputer visionhttps://www.amirrezavishteh.ir/blog/backdoor-detection-lab/https://www.amirrezavishteh.ir/blog/backdoor-detection-lab/A laptop-friendly lab for finetuning a small backdoored model and comparing several detection signals on clean vs. triggered inputs.Sat, 07 Jun 2025 00:00:00 GMTAI safetybackdoorsLLMhttps://www.amirrezavishteh.ir/blog/ftt-nas-review/https://www.amirrezavishteh.ir/blog/ftt-nas-review/An implementation and review of FTT-NAS (Fault-Tolerant Neural Architecture Search), covering the MiBB feature-fault and adSAF weight-fault models for neural networks on unreliable hardware.Sat, 15 Feb 2025 00:00:00 GMTpaper reviewneural architecture searchreliabilityhttps://www.amirrezavishteh.ir/blog/understanding-backdoorbench/https://www.amirrezavishteh.ir/blog/understanding-backdoorbench/A walkthrough of BackdoorBench, the standardized benchmark for evaluating backdoor attacks and defenses in deep learning — what it measures, why it matters, and how to use it.Sat, 15 Feb 2025 00:00:00 GMTAI safetybackdoorsbenchmarkshttps://www.amirrezavishteh.ir/blog/wav2vec2-persian-speech-emotion/https://www.amirrezavishteh.ir/blog/wav2vec2-persian-speech-emotion/Fine-tuning Wav2Vec2 for Persian speech emotion recognition on the ShEMO dataset, combining audio features with text transcripts for robust sentiment classification.Sun, 06 Oct 2024 00:00:00 GMTspeechPersian NLPhttps://www.amirrezavishteh.ir/blog/psychological-health-chatbot/https://www.amirrezavishteh.ir/blog/psychological-health-chatbot/A Persian-language mental-health chatbot built with transformer-based NLP, published at the AbjadNLP workshop — data collection, model training, and evaluation.Fri, 06 Sep 2024 00:00:00 GMTNLPPersian NLPhealthcare AIpublicationhttps://www.amirrezavishteh.ir/blog/osmium-cell-localization/https://www.amirrezavishteh.ir/blog/osmium-cell-localization/An Android app that estimates cellular base-station locations from RSSI and GPS measurements using circular lateration.Mon, 08 Jul 2024 00:00:00 GMTnetworksAndroidcourseworkhttps://www.amirrezavishteh.ir/blog/persian-multimodal-sentiment-analysis/https://www.amirrezavishteh.ir/blog/persian-multimodal-sentiment-analysis/Predicting sentiment from Persian Instagram posts by fusing VGG-16 image features with a bidirectional LSTM over the text.Sat, 10 Feb 2024 00:00:00 GMTNLPmultimodalPersian NLPhttps://www.amirrezavishteh.ir/blog/project-iridium/https://www.amirrezavishteh.ir/blog/project-iridium/A Docker-based network simulation with attacker, victim, and web-server stations: port scanning, credential testing, and data exfiltration in a sandbox.Mon, 04 Dec 2023 00:00:00 GMTsecuritynetworkscourseworkhttps://www.amirrezavishteh.ir/blog/nlp-lab-internship/https://www.amirrezavishteh.ir/blog/nlp-lab-internship/Notes from my research internship at the IUST NLP Lab — the papers I read and the Coursera courses I took along the way.Sat, 04 Nov 2023 00:00:00 GMTNLPresearch