All challenge writeups

El-Gayar Inc.

Engineer network-log features, identify anomalous traffic, and reconstruct the hidden message.

As a senior Analyst, help Mr. Tamer El-Gayar in identifying the hidden message left by the attackers before the operation happens.

Overview

You’re given a network activity dataset (logs.csv) from El-Gayar Inc. The company suspects a stealthy APT attacker has embedded hidden communications in the traffic. Your job is to detect the anomalies, extract the embedded signal, and reconstruct the attacker’s hidden message — the flag.


Phase 1 — Explore the Data & Look for Inconsistencies

What to do (explanation only)

  • Load the CSV (logs.csv) and parse timestamps.
  • Inspect the columns:
    • bytes_up, bytes_down, req_rate, latency_ms, error_rate, packets, etc.
  • Engineer new features to reveal correlations:
    • bytes_ratio = bytes_up / bytes_down
    • bytes_per_packet = (bytes_up + bytes_down) / packets
    • latency_per_req = latency_ms / req_rate

Key Insight: Normal enterprise traffic follows statistical relationships (e.g., more packets → more bytes, higher request rate → lower latency). The anomalies break these patterns.

Code snippet — load & engineer features

import pandas as pd

# Load and sort by time
df = pd.read_csv("logs.csv", parse_dates=["ts"])
df = df.sort_values("ts").reset_index(drop=True)

# Derived metrics
total_bytes = df["bytes_up"] + df["bytes_down"]
bytes_ratio = df["bytes_up"] / (df["bytes_down"] + 1e-9)
bytes_per_packet = total_bytes / (df["packets"] + 1e-9)
latency_per_req = df["latency_ms"] / (df["req_rate"] + 1e-9)

Phase 2 — Detect Anomalies with Machine Learning

What to do (explanation only)

  • Use Isolation Forest (unsupervised anomaly detection).
  • Standardize features before training.
  • Rank rows by anomaly score.

This step surfaces ~60 most suspicious records out of 5,000. The attacker’s activity is hidden among them.

Code snippet — anomaly detection

from sklearn.ensemble import IsolationForest
from sklearn.preprocessing import StandardScaler

# Build feature matrix
X = pd.DataFrame({
    "login_hour": df["login_hour"],
    "bytes_up": df["bytes_up"],
    "bytes_down": df["bytes_down"],
    "req_rate": df["req_rate"],
    "latency_ms": df["latency_ms"],
    "error_rate": df["error_rate"],
    "packets": df["packets"],
    "total_bytes": total_bytes,
    "bytes_ratio": bytes_ratio,
    "bytes_per_packet": bytes_per_packet,
    "latency_per_req": latency_per_req
})

# Standardize & run Isolation Forest
scaler = StandardScaler()
X_scaled = scaler.fit_transform(X)

iforest = IsolationForest(n_estimators=300, contamination=0.01, random_state=1337)
scores = iforest.fit(X_scaled).decision_function(X_scaled)

df["anomaly_score"] = scores
top = df.nsmallest(60, "anomaly_score")

Phase 3 — Identify the Signal

What to do (explanation only)

  • Among anomalies, look at the req_rate column.
  • Normally, req_rate is fractional.
  • Signal rows have exact integer values → these integers are ASCII codes.

Code snippet — extract integer rows

def is_integer_like(x, tol=1e-6):
    return abs(x - round(x)) < tol

signal_rows = top[top["req_rate"].apply(is_integer_like)]

Phase 4 — Reconstruct the Hidden Message

What to do (explanation only)

  • Sort anomalous rows by timestamp.
  • Convert each req_rate integer to its ASCII character.
  • Concatenate to form the flag.

Code snippet — decode ASCII

signal_rows = signal_rows.sort_values("ts")
flag = "".join(chr(int(round(v))) for v in signal_rows["req_rate"])
print("Flag:", flag)
  • Final Flag: IEEE{twenty_low_chr}

Original Notion writeup