TTS-Arena-V2 / security.py
GitHub Actions
Sync from GitHub repo
5259aa6
raw
history blame
12.7 kB
"""
Security utilities for TTS Arena to prevent vote manipulation and botting.
"""
from datetime import datetime, timedelta
from models import db, Vote, User
from sqlalchemy import func, and_, or_
import logging
logger = logging.getLogger(__name__)
def detect_suspicious_voting_patterns(user_id, hours_back=24, max_votes_per_hour=30):
"""
Detect if a user has suspicious voting patterns.
Updated to allow rapid voting for reasonable periods (30 votes/hour = 1 vote every 2 minutes)
Returns (is_suspicious, reason, vote_count)
"""
if not user_id:
return False, None, 0
# Check voting frequency over 24 hours
time_threshold = datetime.utcnow() - timedelta(hours=hours_back)
recent_votes = Vote.query.filter(
and_(
Vote.user_id == user_id,
Vote.vote_date >= time_threshold
)
).count()
# Allow up to 30 votes per hour (720 votes in 24 hours)
# This allows rapid voting for several hours but catches extended botting
max_votes_24h = max_votes_per_hour * hours_back
if recent_votes > max_votes_24h:
return True, f"Too many votes: {recent_votes} in {hours_back} hours (max: {max_votes_24h})", recent_votes
# Additional check: if someone votes more than 100 times in 3 hours, that's suspicious
# (100 votes in 3 hours = 1 vote every 1.8 minutes, which is very sustained)
if hours_back >= 3:
three_hour_threshold = datetime.utcnow() - timedelta(hours=3)
votes_3h = Vote.query.filter(
and_(
Vote.user_id == user_id,
Vote.vote_date >= three_hour_threshold
)
).count()
if votes_3h > 100:
return True, f"Excessive voting in short period: {votes_3h} votes in 3 hours", recent_votes
return False, None, recent_votes
def detect_model_bias(user_id, model_id, min_votes=5, bias_threshold=0.8):
"""
Detect if a user consistently votes for a specific model.
Returns (is_biased, bias_ratio, total_votes_for_model, total_votes)
"""
if not user_id:
return False, 0, 0, 0
# Get all votes by this user
total_votes = Vote.query.filter_by(user_id=user_id).count()
if total_votes < min_votes:
return False, 0, 0, total_votes
# Get votes where this user chose the specific model
votes_for_model = Vote.query.filter(
and_(
Vote.user_id == user_id,
Vote.model_chosen == model_id
)
).count()
bias_ratio = votes_for_model / total_votes if total_votes > 0 else 0
is_biased = bias_ratio >= bias_threshold and total_votes >= min_votes
return is_biased, bias_ratio, votes_for_model, total_votes
def detect_coordinated_voting(model_id, hours_back=6, min_users=3, vote_threshold=10):
"""
Detect coordinated voting campaigns for a specific model.
Returns (is_coordinated, user_count, vote_count, suspicious_users)
"""
time_threshold = datetime.utcnow() - timedelta(hours=hours_back)
# Get recent votes for this model
recent_votes = db.session.query(Vote.user_id).filter(
and_(
Vote.model_chosen == model_id,
Vote.vote_date >= time_threshold
)
).all()
if len(recent_votes) < vote_threshold:
return False, 0, len(recent_votes), []
# Count unique users
unique_users = set(vote.user_id for vote in recent_votes if vote.user_id)
user_count = len(unique_users)
# Check if multiple users are voting for the same model in a short time
if user_count >= min_users and len(recent_votes) >= vote_threshold:
# Get user details for suspicious users
suspicious_users = []
for user_id in unique_users:
user_votes_for_model = Vote.query.filter(
and_(
Vote.user_id == user_id,
Vote.model_chosen == model_id,
Vote.vote_date >= time_threshold
)
).count()
if user_votes_for_model > 1: # Multiple votes for same model in short time
user = User.query.get(user_id)
if user:
suspicious_users.append({
'user_id': user_id,
'username': user.username,
'votes_for_model': user_votes_for_model,
'account_age_days': (datetime.utcnow() - user.join_date).days if user.join_date else None
})
return True, user_count, len(recent_votes), suspicious_users
return False, user_count, len(recent_votes), []
def detect_rapid_voting(user_id, min_interval_seconds=3):
"""
Detect if a user is voting too rapidly (potential bot behavior).
This allows rapid voting (3+ seconds) for reasonable periods, but flags
extended periods of very rapid voting that indicate bot behavior.
Returns (is_rapid, intervals, avg_interval)
"""
if not user_id:
return False, [], 0
# Get more recent votes to better analyze patterns (last 50 instead of 10)
recent_votes = Vote.query.filter_by(user_id=user_id).order_by(
Vote.vote_date.desc()
).limit(50).all()
if len(recent_votes) < 50: # Need at least 50 votes to detect patterns
return False, [], 0
# Calculate intervals between votes
intervals = []
for i in range(len(recent_votes) - 1):
interval = (recent_votes[i].vote_date - recent_votes[i + 1].vote_date).total_seconds()
intervals.append(interval)
avg_interval = sum(intervals) / len(intervals) if intervals else 0
# More sophisticated bot detection:
# 1. Count votes with intervals < 3 seconds (very rapid)
very_rapid_votes = sum(1 for interval in intervals if interval < 3)
# 2. Count votes with intervals < 1 second (extremely rapid - likely bot)
extremely_rapid_votes = sum(1 for interval in intervals if interval < 1)
# 3. Check for sustained rapid voting patterns
# Look for sequences of 10+ votes all under 5 seconds
sustained_rapid_sequences = 0
current_sequence = 0
for interval in intervals:
if interval < 5:
current_sequence += 1
else:
if current_sequence >= 10: # 10+ votes in a row under 5 seconds
sustained_rapid_sequences += 1
current_sequence = 0
# Final check for remaining sequence
if current_sequence >= 10:
sustained_rapid_sequences += 1
# Flag as rapid/bot if:
# - More than 20% of votes are extremely rapid (< 1 second) OR
# - More than 60% of votes are very rapid (< 3 seconds) AND there are sustained sequences OR
# - There are multiple sustained rapid sequences (10+ votes under 5 seconds each)
total_intervals = len(intervals)
extremely_rapid_ratio = extremely_rapid_votes / total_intervals if total_intervals > 0 else 0
very_rapid_ratio = very_rapid_votes / total_intervals if total_intervals > 0 else 0
is_rapid = (
extremely_rapid_ratio > 0.2 or # > 20% extremely rapid
(very_rapid_ratio > 0.6 and sustained_rapid_sequences > 0) or # > 60% very rapid + sustained
sustained_rapid_sequences >= 2 # Multiple sustained rapid sequences
)
return is_rapid, intervals, avg_interval
def check_user_security_score(user_id):
"""
Calculate a security score for a user based on various factors.
Returns (score, factors) where score is 0-100 (higher = more trustworthy)
"""
if not user_id:
return 0, {"error": "No user ID provided"}
user = User.query.get(user_id)
if not user:
return 0, {"error": "User not found"}
factors = {}
score = 100 # Start with perfect score and deduct points
# Account age factor
if user.join_date:
account_age_days = (datetime.utcnow() - user.join_date).days
factors['account_age_days'] = account_age_days
if account_age_days < 45:
score -= 30
elif account_age_days < 90:
score -= 15
elif account_age_days < 180:
score -= 5
else:
score -= 20
factors['account_age_days'] = None
# HF account age factor
if user.hf_account_created:
hf_age_days = (datetime.utcnow() - user.hf_account_created).days
factors['hf_account_age_days'] = hf_age_days
if hf_age_days < 30:
score -= 25 # This should be caught by auth, but double-check
elif hf_age_days < 90:
score -= 10
else:
score -= 15
factors['hf_account_age_days'] = None
# Voting pattern analysis
is_suspicious, reason, vote_count = detect_suspicious_voting_patterns(user_id)
factors['suspicious_voting'] = is_suspicious
factors['recent_vote_count'] = vote_count
if is_suspicious:
score -= 25
factors['suspicious_reason'] = reason
# Rapid voting check
is_rapid, intervals, avg_interval = detect_rapid_voting(user_id)
factors['rapid_voting'] = is_rapid
factors['avg_vote_interval'] = avg_interval
if is_rapid:
score -= 20
# Total vote count (very new users with many votes are suspicious)
total_votes = Vote.query.filter_by(user_id=user_id).count()
factors['total_votes'] = total_votes
if account_age_days and account_age_days < 7 and total_votes > 20:
score -= 15 # New account with many votes
# Model bias detection - check for extreme bias toward any single model
if total_votes >= 5: # Only check if user has enough votes
max_bias_ratio = 0
most_biased_model = None
# Get all models this user has voted for
user_votes = Vote.query.filter_by(user_id=user_id).all()
model_stats = {}
for vote in user_votes:
chosen_id = vote.model_chosen
rejected_id = vote.model_rejected
# Track appearances and choices
if chosen_id not in model_stats:
model_stats[chosen_id] = {'chosen': 0, 'appeared': 0}
if rejected_id not in model_stats:
model_stats[rejected_id] = {'chosen': 0, 'appeared': 0}
model_stats[chosen_id]['chosen'] += 1
model_stats[chosen_id]['appeared'] += 1
model_stats[rejected_id]['appeared'] += 1
# Find the highest bias ratio
for model_id, stats in model_stats.items():
if stats['appeared'] >= 5: # Only consider models with enough appearances
bias_ratio = stats['chosen'] / stats['appeared']
if bias_ratio > max_bias_ratio:
max_bias_ratio = bias_ratio
most_biased_model = model_id
factors['max_bias_ratio'] = max_bias_ratio
factors['most_biased_model_id'] = most_biased_model
# Deduct points based on bias level
if max_bias_ratio >= 0.95: # 95%+ bias
score -= 30
factors['bias_penalty'] = 'Extreme bias (95%+)'
elif max_bias_ratio >= 0.9: # 90%+ bias
score -= 20
factors['bias_penalty'] = 'Very high bias (90%+)'
elif max_bias_ratio >= 0.8: # 80%+ bias
score -= 10
factors['bias_penalty'] = 'High bias (80%+)'
else:
factors['bias_penalty'] = None
else:
factors['max_bias_ratio'] = 0
factors['bias_penalty'] = None
# Ensure score doesn't go below 0
score = max(0, score)
factors['final_score'] = score
return score, factors
def is_vote_allowed(user_id, ip_address=None):
"""
Check if a vote should be allowed based on security factors.
Returns (allowed, reason, security_score)
"""
if not user_id:
return False, "User not authenticated", 0
# Check security score
score, factors = check_user_security_score(user_id)
# Very low scores are blocked
if score < 20:
return False, f"Security score too low: {score}/100", score
# Check for recent suspicious activity
if factors.get('suspicious_voting'):
return False, f"Suspicious voting pattern detected: {factors.get('suspicious_reason')}", score
if factors.get('rapid_voting'):
return False, f"Voting too rapidly (avg interval: {factors.get('avg_vote_interval', 0):.1f}s)", score
# Additional IP-based checks could go here
return True, "Vote allowed", score