SafeScreen: A Safety-First Screening Framework for Personalized Video Retrieval for Vulnerable Users

Wenzheng Zhao; Madhava Kalyan Gadiputi; Fengpei Yuan

cs.CVMar 12, 2026

SafeScreen: A Safety-First Screening Framework for Personalized Video Retrieval for Vulnerable Users

Authors: Wenzheng Zhao, Madhava Kalyan Gadiputi, Fengpei Yuan

Organizations: Worcester Polytechnic Institute, Worcester, MA, USA

Abstract

Open-domain video platforms offer rich, personalized content that could support health, caregiving, and educational applications, but their engagement-optimized recommendation algorithms can expose vulnerable users to inappropriate or harmful material. These risks are especially acute in child-directed and care settings (e.g., dementia care), where content must satisfy individualized safety constraints before being shown. We introduce SafeScreen, a safety-first video screening framework that retrieves and presents personalized video while enforcing individualized safety constraints. Rather than ranking videos by relevance or popularity, SafeScreen treats safety as a prerequisite and performs sequential approval or rejection of candidate videos through an automated pipeline. SafeScreen integrates three key components: (i) profile-driven extraction of individualized safety criteria, (ii) evidence-grounded assessments via adaptive question generation and multimodal VideoRAG analysis, and (iii) LLM-based decision-making that verifies safety, appropriateness, and relevance before content exposure. This design enables explainable, real-time screening of uncurated video repositories without relying on precomputed safety labels. We evaluate SafeScreen in a dementia-care reminiscence case study using 30 synthetic patient profiles and 90 test queries. Results demonstrate that SafeScreen prioritizes safety over engagement, diverging from YouTube's engagement-optimized rankings in 80-93% of cases, while maintaining high levels of safety coverage, sensibleness, and groundedness, as validated by both LLM-based evaluation and domain experts.

Figures & tables

Figure 1. Conceptual comparison between conventional engagement-driven video recommendation systems (top) and SafeScreen, a safety-first personalization framework for reminiscence video retrieval (bottom). SafeScreen reverses the standard optimization objective by prioritizing individual safety constraints and multimodal verification over popularity or crowd-based similarity. Overview

Figure 2. Complete SafeScreen framework overview showing the three-stage pipeline: (1) Stage 1 (green): Prefiltering steps including risk detection, risk-aware profile extraction, preference extraction, and candidate video retrieval. The system requests permission for medium/high-risk queries or terminates if permission is denied. (2) Stage 2 (orange): VideoRAG Analysis where an LLM generates patient-specific safety questions based on the extracted profile and query, then VideoRAG analyzes candidate videos to produce evidence-grounded Q/A pairs. (3) Stage 3 (purple): LLM Evaluation implementing sequential safety screening. Videos are evaluated one at a time; the first to pass all safety criteria is selected, while failed videos are rejected immediately. The process continues until an acceptable video is found or all candidates are exhausted. User inputs (cyan) flow through sequential safety verification before video selection.

Panel of Figure 3 — Figure 3. SafeScreen deployment contexts: clinical integration (left) and systematic evaluation (right).

Profile Component	Data Elements
Demographics	Age, diagnosis, cognitive stage (mild/moderate/severe)
Personal History	Past occupation, era preferences, cultural background
Interests	Topics of engagement (e.g., music, nature, sports)
Sensitivities	Auditory triggers (sirens, loud noises), visual triggers (medical imagery, flashing lights), content triggers (loss themes, emergency situations)
Cognitive Characteristics	Attention span, complexity tolerance, communication needs, preferred pacing
Engagement History	Previously successful content, previously distressing content

Table 1. Profile Schema for Dementia Patient Population

Component	Input	Output
Risk Detection	User query, population type, cognitive level	Risk level (LOW/MEDIUM/HIGH), permission required (yes/no), reasoning
Profile Extraction	User profile, query, context	Safety constraints (triggers to avoid), engagement factors (interests, era), appropriateness requirements (complexity, pacing), relevance criteria
Candidate Retrieval	Refined query	3-5 candidate videos with metadata (ID, title, URL, duration, channel)

Table 2. Stage 1: Prefiltering Data Flow

Component	Input	Output
Question Generation	Extracted criteria, user query	5-6 adaptive safety questions with purpose labels (safety check, relevance, appropriateness)
VideoRAG Analysis	Generated questions, video segment (first 5 min)	Answers with detailed observations, evidence timestamps (MM:SS-MM:SS), confidence levels (high/medium/low)

Table 3. Stage 2: VideoRAG Analysis Data Flow

Component	Input	Output
Safety Evaluation	VideoRAG Q/A pairs, patient profile	Decision (APPROVE/REJECT), confidence level, safety verification results per category (auditory, visual, cognitive, relevance), therapeutic value assessment
Sequential Screening	Evaluation results	If APPROVE: selected video with evidence summary, personalization notes, caregiver/guardian guidance. If REJECT: proceed to next candidate

Table 4. Stage 3: LLM Evaluation Data Flow

Single-video cases (first video passed)
Scenario	Level 1	Level 2	Level 3
Match with YouTube top result	23.1%	25.0%	11.1%
Overall	20.6%
Multi-video cases (2-3 videos screened)
Match with YouTube top result	0%	0%	14.3%
Overall	7.1%

Table 5. Comparison of SafeScreen selections vs. YouTube top search results

Metric	Score (1-5 scale)
Safety Coverage	4.26 $\pm$ 1.09
Sensibleness	4.42 $\pm$ 0.95
Groundedness	4.26 $\pm$ 1.16

Table 6. LLM evaluation across 90 test cases (mean

\pm

SD)

Metric	Expert Mean	Weighed Cohen’s $\kappa$
Safety Coverage	4.5	0.88
Sensibleness	3.5	0.92
Groundedness	4	0.94
Overall Agreement		Substantial

Table 7. Expert validation scores on stratified 20% samples

Dementia Care

See all 18 papers

SafeLens: Deliberate and Efficient Video Guardrails with Fast-and-Slow Screening
May 17, 2026Shahriar Kabir Nahin, Hadi Askari, Muhao Chen +1Adaptive Inference Efficient VLM Inference
University of South Florida · University of California, Davis
The Temporal Moderation Gap: Text-to-Video Safety Filters Are Blind to Harm in Motion
Aug 19, 2026Yuxin Cao, Fusen Guo, Yuezhong Wu +2Video Content Moderation Multimodal Content Moderation
University of New South Wales Australia · Fuzhou University China
SafeCRS: Personalized Safety Alignment for LLM-Based Conversational Recommender Systems
Mar 3, 2026Haochang Hao, Yifan Xu, Xinzhuo Li +2Recommender Systems Large Language Model-Based Recommendation
University of Illinois at Chicago Chicago, IL, USA · University of Notre Dame South Bend, IN, USA · Amazon Santa Clara, CA, USA

May 17, 2026J/K move · Enter open · S save

SafeLens: Deliberate and Efficient Video Guardrails with Fast-and-Slow Screening

Shahriar Kabir Nahin, Hadi Askari, Muhao Chen, Anshuman Chhabra

University of South Florida · University of California, Davis

Adaptive Inference Efficient VLM Inference Video Content Moderation Content Moderation Safety Filtering

Open paper AskPDF arXiv

The rapid growth of online video platforms and AI-generated content has made reliable video guardrails a key challenge for safety and real-world deployment. While most videos can be screened through fast pattern recognition, a small subset requires deeper reasoning over temporally complex content and nuanced policy constraints. Existing approaches typically rely on large vision-language models applied uniformly across all inputs, resulting in high inference costs and inefficient allocation of computation. We propose SafeLens, a video guardrail framework that introduces a fast-and-slow inference architecture for efficient and accurate content moderation with variable computational cost across inputs. Additionally, we construct a high-quality dataset by applying influence-guided filtering to the SafeWatch Dataset, retaining only 2.4% of the original data. To further address limitations of training-time scaling, we enable test-time reasoning by augmenting the filtered data with structured Chain-of-Thought traces. Across real-world and AI-generated video benchmarks, SafeLens achieves state-of-the-art performance, outperforming strong open-source video guardrails (e.g., SafeWatch-8B, OmniGuard-7B) and closed-source models (e.g., GPT-5.4, Gemini-3.1-pro) while significantly reducing inference cost, demonstrating that efficient design serves to be more effective than scaling data or model size alone.

Abstract

Dementia Care

Explore similar work

SafeLens: Deliberate and Efficient Video Guardrails with Fast-and-Slow Screening

The Temporal Moderation Gap: Text-to-Video Safety Filters Are Blind to Harm in Motion

SafeCRS: Personalized Safety Alignment for LLM-Based Conversational Recommender Systems