Senior Design Project (CSE499A & CSE499B)
Department of Electrical & Computer Engineering
North South University, Bangladesh
Spring 2025 (Data Collection) + Summer 2025 (System Development)
IdentiFind is a comprehensive two-phase research project that addresses the challenging problem of person profile disambiguation and biographical generation for individuals with limited online presence. The project evolved through two semesters of intensive work:
CSE499A (Spring 2025) → CSE499B (Summer 2025)
───────────────────── ─────────────────────
Data Collection System Development
Web Scraping (32K+ entries) Face Recognition Pipeline
Dataset Validation (9.4K profiles) NLP & LLM Integration
GPT-4o-mini Summaries Multi-Source Validation
Comprehensive Evaluation
Traditional multimodal entity linking systems face critical limitations:
- Heavy reliance on knowledge bases that often lack information about ordinary individuals
- Vulnerability to noisy, misleading, or contradictory information across web sources
- Inability to handle zero-shot scenarios for unknown entities
- Poor performance on individuals with weak digital footprints
- Existing search engines (Google Lens) require clear, individual portraits and fail with group photos
IdentiFind provides an end-to-end solution requiring only:
- ✅ A single input image (individual portrait or group photo)
- ✅ Unlabeled web links (automatically retrieved via FaceCheck API)
Key Innovation: Zero-shot biographical profiling without pre-existing knowledge bases, achieving high-confidence profiles through multi-source validation and Natural Language Inference.
| Name | Student ID |
|---|---|
| Tamim Ishrak Sanjid | 2122405642 |
| Md. Shakib Shahariar Junayed | 2121095042 |
| MD. Sakib Sami | 2121422042 |
| Md. Fahim Morshed | 2122279642 |
Faculty Advisor: Dr. Nabeel Mohammed
Associate Professor, Department of Electrical & Computer Engineering
Research Assistants: Md. Mohibur Rahman Nabil, Afsara Waziha
Face2Profile Benchmark Evaluation:
| Metric | Mistral-7B | DeepSeek-7B |
|---|---|---|
| Total Profiles | 10,000 | 10,000 |
| Successfully Processed | 3,474 (34.7%) | 3,474 (34.7%) |
| Failed (Anti-scraping) | 6,526 (65.3%) | 6,526 (65.3%) |
| Avg. Processing Time | 3.2 min/profile | 3.5 min/profile |
| Total Processing Time | 185 hours | 202 hours |
Individual Component Matching:
| Match Type | Mistral-7B | DeepSeek-7B | Difference |
|---|---|---|---|
| Full Name | 2,329 (67.0%) | 2,274 (65.5%) | +55 (+2.4%) |
| First Name | 3,130 (90.1%) | 3,066 (88.3%) | +64 (+2.1%) |
| Last Name | 3,020 (86.9%) | 3,062 (88.1%) | -42 (-1.4%) |
| First + Last | 2,440 (70.2%) | 2,392 (68.8%) | +48 (+2.0%) |
Professional Information Extraction:
| Match Type | Mistral-7B | DeepSeek-7B | Difference | % Improvement |
|---|---|---|---|---|
| Name + Company | 1,558 (44.8%) | 1,483 (42.7%) | +75 | +5.1% |
| Name + Job | 957 (27.5%) | 1,383 (39.8%) | -426 | +44.5% ⭐ |
| Name + Company + Job | 433 (12.5%) | 606 (17.4%) | -173 | +40.0% ⭐ |
Key Finding: DeepSeek-7B significantly outperforms Mistral-7B in professional context extraction
| Metric Category | Metric | Mistral-7B | DeepSeek-7B | Winner |
|---|---|---|---|---|
| Lexical Quality | ROUGE-1 | 0.500 ⭐ | 0.472 | Mistral |
| ROUGE-2 | 0.243 ⭐ | 0.197 | Mistral | |
| ROUGE-L | 0.315 ⭐ | 0.272 | Mistral | |
| Jaccard Similarity | 0.312 ⭐ | 0.272 | Mistral | |
| METEOR | 0.359 ⭐ | 0.337 | Mistral | |
| Semantic | BERTScore | 0.897 | 0.901 ⭐ | DeepSeek |
| Length Ratio | 0.732 | 0.731 | Tie | |
| Precision (BLEU) | BLEU-1 | 0.347 ⭐ | 0.323 | Mistral |
| BLEU-2 | 0.160 ⭐ | 0.123 | Mistral | |
| BLEU-4 | 0.068 ⭐ | 0.043 | Mistral | |
| C-BLEU-1 | 0.313 ⭐ | 0.295 | Mistral | |
| C-BLEU-2 | 0.145 ⭐ | 0.113 | Mistral | |
| C-BLEU-4 | 0.062 ⭐ | 0.039 | Mistral | |
| Entity Coverage | ECS_overall | 0.467 | 0.530 ⭐ | DeepSeek |
| ECS_noname | 0.290 | 0.425 ⭐ | DeepSeek | |
| ECS_nojob | 0.645 ⭐ | 0.634 | Mistral | |
| Diversity | Bigram Diversity | 0.981 | 0.994 ⭐ | DeepSeek |
| Trigram Diversity | 0.992 | 0.995 ⭐ | DeepSeek | |
| Semantic Diversity | 1.075 ⭐ | 1.044 | Mistral | |
| Vocab Overlap | 0.263 ⭐ | 0.217 | Mistral |
Mistral-7B Strengths: ✅
- Superior linguistic coherence (ROUGE-1: 0.500 vs 0.472)
- Better basic name extraction (67.0% vs 65.5%)
- Higher lexical precision across all BLEU metrics (58% better BLEU-4)
- Enhanced semantic diversity (1.075 vs 1.044)
- Better vocabulary richness (26.3% vs 21.7% overlap)
DeepSeek-7B Strengths: ✅
- Exceptional professional info extraction (44.5% better name+job matching)
- Superior job title identification (46.6% improvement in ECS_noname)
- Higher overall entity coverage (ECS: 0.530 vs 0.467 = +13.5%)
- Better complex entity combinations (40.0% improvement in name+company+job)
- Stronger semantic understanding (BERTScore: 0.901 vs 0.897)
Entity Extraction Performance:
Name + Job Title Extraction
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Mistral-7B: ████████████░░░░░░░░░░░░░░░░░░ 27.5% (957)
DeepSeek-7B: █████████████████░░░░░░░░░░░░░ 39.8% (1,383) ⭐
Improvement: +44.5%
Lexical Quality (ROUGE-1):
ROUGE-1 Score (Higher = Better)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Mistral-7B: ██████████████████████████░░░░ 0.500 ⭐
DeepSeek-7B: ████████████████████████░░░░░░ 0.472
Difference: +5.9%
Overall Entity Coverage (ECS):
Entity Coverage Score
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Mistral-7B: ███████████████████░░░░░░░░░░░ 0.467
DeepSeek-7B: █████████████████████░░░░░░░░░ 0.530 ⭐
Improvement: +13.5%
Decision Matrix: When to Use Each Model
| Use Case | Recommended Model | Reason |
|---|---|---|
| Professional recruitment | DeepSeek-7B | Superior job title + company extraction |
| Biography writing | Mistral-7B | Better linguistic quality and coherence |
| LinkedIn profile generation | DeepSeek-7B | Excellent professional context understanding |
| Academic profiles | Mistral-7B | More fluent summaries |
| Legal/medical verification | DeepSeek-7B | Higher factual accuracy (ECS) |
| Creative writing | Mistral-7B | Better vocabulary diversity |
| Data extraction tasks | DeepSeek-7B | Superior entity extraction |
| Public communication | Mistral-7B | More readable output |
IdentiFind processes publicly available data only. However, users must be aware of ethical responsibilities:
✅ Acceptable Use:
- Legitimate identity verification (with consent)
- Missing person identification (law enforcement)
- Academic research (IRB approved)
- Professional background checks (with authorization)
❌ Prohibited Use:
- Unauthorized surveillance or stalking
- Discriminatory profiling or hiring practices
- Invasion of privacy without consent
- Malicious impersonation or identity theft
CSE499A Dataset (9,472 profiles):
- ✅ Collected from publicly accessible professional platforms
- ✅ No scraping of private social media (Facebook, Instagram)
- ✅ Respect for robots.txt and platform ToS
- ✅ No storage of sensitive PII (SSN, credit cards, etc.)
We acknowledge and actively work to mitigate biases in:
-
Facial Recognition Systems
- Demographic disparities (see Limitations section)
- Age bias (better for 25-55 age range)
- Ongoing bias auditing framework development
-
Web Data Availability
- Digital divide (well-documented individuals favored)
- Geographic bias (US profiles more accessible)
- Language bias (English-centric processing)
-
Language Processing
- English-only support (planned multilingual expansion)
- Cultural context understanding
- Name recognition across cultures
We commit to:
- 🔍 Open-source codebase for community review
- 📊 Public bias audit reports (quarterly)
- 📝 Detailed methodology documentation
- 🔐 Security vulnerability disclosure program
Users have the right to:
- 🚫 Opt-out of profiling (if detected)
- 📜 Request explanation of generated profile
- ✏️ Correct inaccurate information
- 🗑️ Request profile deletion
This project aligns with:
SDG #9: Industry, Innovation and Infrastructure
- Innovation in AI-powered identity verification
- Accessible technology for ordinary individuals
- Scalable infrastructure for information processing
SDG #16: Peace, Justice and Strong Institutions
- Legitimate identity verification for legal systems
- Missing person identification capabilities
- Combating misinformation through source validation