Bird
Raised Fist0

Spam Detection System - Features, Models, Adversarial Robustness

SYSTEM-DESIGNmediumInterview freq: medium
Choose your preparation mode6 modes available
LearnTraceDebugTrade-offChallenge
🤖
Spam Detection System - Features, Models, Adversarial Robustness
Design a robust spam detection system with adversarial resilience.
system-designmediumFreq: mediumTrust & Safety
Prerequisites:Hate Speech DetectionModel Debugging
🎯
Problem Framing
Business Objective

Minimize spam content exposure to users to enhance user trust and platform integrity.

ML Objective

Binary classification to accurately identify spam messages with high precision and recall.

Constraints
  • Latency SLA: P99 < 100ms for real-time detection
  • Label availability: Limited labeled data for new spam types
  • Adversarial robustness: System must withstand adversarial attacks
  • Regulatory compliance: Must comply with data privacy laws
  • Cold start: Handle new users and content types effectively
Clarifying Questions
  • What is the current false positive rate for spam detection?
  • How frequently do new types of spam emerge?
  • What are the regulatory requirements for data handling?
  • What is the expected volume of messages per second?
  • Are there existing rule-based systems in place?
Out of Scope
✗ Non-textual spam detection (e.g., images, videos)
✗ User interface design
✗ Non-English language support
📊
Offline Metrics
Precision: To minimize false positives, ensuring legitimate messages are not flagged as spam.
Recall: To maximize the detection of spam messages, ensuring minimal spam exposure.
📈
Online Metrics
False Positive Rate: Percentage of legitimate messages incorrectly flagged as spam.
User Report Rate: Number of spam messages reported by users post-detection.
⚖️
Trade-offs

High precision may reduce recall, leading to more spam exposure. Balancing precision and recall is crucial.

💼
Business Metric

User trust score, reflecting user satisfaction and platform integrity.

🗄️
Data Sources
  • User messages: Text data, real-time ingestion
  • User reports: Feedback data, daily batch updates
  • Historical spam data: Labeled dataset, weekly updates
🏷️
Data Properties
Labeling Strategy

Programmatic labeling using user reports and manual review for edge cases.

Data Volume

Approximately 10 million messages per day, 500GB/month

Freshness

Features must be updated in real-time to capture the latest spam patterns.

⚠️
Data Challenges
  • Sparse labels for new spam types
  • Adversarial examples crafted to bypass detection
  • Cold start for new users and content types
  • Data privacy and compliance issues
🔧
Key Features
numerical
Message Length
Spam messages often have characteristic lengths.
numerical
Keyword Frequency
Certain keywords are indicative of spam.
categorical
Sender Reputation
Known spammers can be identified by reputation.
numerical
URL Count
Spam messages often contain multiple URLs.
categorical
Time of Day
Spam activity may peak at certain times.
embedding
Message Embedding
Captures semantic meaning of the message.
⚡
Feature Engineering Notes
Real-time vs Batch

Real-time computation for message length, keyword frequency, and URL count; batch updates for sender reputation.

Feature Store

Required

🐍
Feature Snippet
Python
def compute_keyword_frequency(message, keywords):
    return sum(1 for word in message.split() if word in keywords) / len(message.split())
🤖
Model Selection
Baseline

Logistic Regression with TF-IDF features

Proposed Model

BERT-based transformer model

BERT captures contextual nuances and is robust against adversarial text manipulations.
✗
Alternatives Rejected
Random Forest: Rejected due to inability to handle high-dimensional text data effectively.
XGBoost: Rejected due to latency constraints and lack of contextual understanding.
🏋️
Training Strategy

Offline training with periodic updates using new labeled data.

🚀
Serving Infrastructure
LatencyP99 < 100ms end-to-end
Throughput10,000 QPS
ArchitectureSingle-stage model serving for simplicity and speed.
CachingCache recent predictions with a TTL of 5 minutes to reduce load.
FallbackRule-based system as a fallback when the model is unavailable.
📡
Data Drift Signals
Shift in message length distribution
Change in keyword frequency patterns
New URL patterns emerging
📉
Model Drift Signals
Decrease in precision over time
Increase in false positive rate
Shift in prediction score distribution
🔔
Alerting & Retraining
Retraining Trigger

Precision drops below 90% or false positive rate exceeds 5%.

Alerting Strategy

On-call engineer paged for precision drop, logs for false positives, auto-recovery for minor drifts.

🎯
Interview Angles
  • How to ensure adversarial robustness in spam detection?
  • Balancing precision and recall in a spam detection system.
  • Handling data privacy and compliance in spam detection.
⚠️
Common Mistakes
Ignoring adversarial robustness in model selection.
Overfitting to historical spam patterns without considering new types.
Neglecting real-time feature computation needs.
Failing to define clear retraining triggers.
📌
Key Decisions
  • Choosing BERT for its contextual understanding and adversarial robustness.
  • Implementing a real-time feature computation pipeline.
  • Defining clear metrics and retraining triggers to maintain model performance.

Practice

(1/5)
1. For a content-based recommendation system, which metric is most appropriate to evaluate the quality of item embeddings?
easy
A. Cosine similarity
B. Precision
C. Recall
D. Mean Absolute Error (MAE)

Solution

  1. Step 1: Understand the evaluation context

    Item embeddings are evaluated based on how well they capture similarities, which is best measured by cosine similarity.

  2. Step 2: Eliminate wrong options

    A & B: Precision and Recall are more suited for classification tasks. D: MAE is used for regression, not similarity.

  3. Final Answer:

    Option A -> Option A

  4. Quick Check:

    Cosine similarity measures embedding quality. [OK]
Hint: Cosine similarity measures embedding quality. [OK]
Common Mistakes:
2. In the context of avoiding filter bubbles in recommendation systems, which component is responsible for ensuring content diversity?
easy
A. Collaborative filtering
B. Diversity re-ranking layer
C. User profiling
D. Content-based filtering

Solution

  1. Step 1: Identify the role of each component

    Collaborative and content-based filtering focus on user preferences, while diversity re-ranking ensures varied content.

  2. Step 2: Eliminate wrong options

    Collaborative filtering (A) and content-based filtering (B) do not inherently ensure diversity. User profiling (D) focuses on understanding user preferences, not diversity.

  3. Final Answer:

    Option B -> Option B

  4. Quick Check:

    Diversity re-ranking ensures varied content [OK]
Hint: Diversity re-ranking ensures varied content [OK]
Common Mistakes:
3. In an A/B testing platform, what is a likely failure mode if the randomization engine is not functioning correctly?
medium
A. The experiment results are biased.
B. Users are not consistently assigned to the same variant.
C. The platform cannot track user interactions.
D. The experiment dashboard fails to update.

Solution

  1. Step 1: Identify the role of the randomization engine.

    It ensures unbiased assignment of users to variants.

  2. Step 2: Eliminate wrong options.

    Session persistence handles consistent assignment, tracking is separate, and dashboard updates are unrelated.

  3. Final Answer:

    Option A -> Option A

  4. Quick Check:

    Randomization ensures unbiased results. [OK]
Hint: Randomization ensures unbiased results. [OK]
Common Mistakes:
4. In a pairwise learning to rank system, what is a common failure mode if the model is not performing well?
medium
A. The model is overfitting due to too many features
B. The model is not capturing the relative order of item pairs
C. The model is not considering the entire list of items
D. The model is using an inappropriate loss function for regression

Solution

  1. Step 1: Identify the failure mode specific to pairwise ranking.

    Pairwise models can fail if they don't capture the relative order of item pairs.

  2. Step 2: Eliminate wrong options.

    Overfitting and inappropriate loss functions are generic issues, and pairwise doesn't consider the entire list, making C incorrect.

  3. Final Answer:

    Option B -> Option B

  4. Quick Check:

    Pairwise = focus on item pairs [OK]
Hint: Pairwise = focus on item pairs [OK]
Common Mistakes:
5. When deploying a Two-Tower Model for Retrieval, which approach is better for handling cold-start users and why?
medium
A. Use random embeddings until enough data is collected.
B. Use collaborative filtering to generate initial embeddings.
C. Use content-based features to generate initial embeddings.
D. Use embeddings from a pre-trained language model.

Solution

  1. Step 1: Consider the cold-start problem

    Cold-start users lack interaction data, making content-based features useful for initial embeddings.

  2. Step 2: Eliminate wrong options

    Use collaborative filtering to generate initial embeddings. relies on interaction data, which cold-start users lack. Use random embeddings until enough data is collected. provides no meaningful information. Use embeddings from a pre-trained language model. might not capture user-specific features effectively.

  3. Final Answer:

    Option C -> Option C

  4. Quick Check:

    Content-based features help with cold-start. [OK]
Hint: Content-based features help with cold-start. [OK]
Common Mistakes:
Recommended

Ready to prove it?

Pass the test and this pattern moves unknown strong while earning +1 Grip toward your certificate.

10 questions~5 mineasy hard