Radio Frequency Rumble
Inspect a Random Forest, recover its hidden class, then decode a message concealed in two joined datasets.
Someone’s been broadcasting secret messages, but their code is more scrambled than your breakfast eggs. Can you tune in, decode the madness, and claim your flag before the static drives you nuts?
Overview
Files provided:
- rf_100_classes2.joblib
which is a machine learning model file. The used model was Random Forest Classifier, which classifies several classes based on an input.
The input consists of an array with 6 numerical elements. The goal here is to reach the flag, but the intended approach is split into 2 phases
Phase 1 Solution
- The first part of the intended solution approach is to load the model into a python environment and use it. It takes in an array of 6 elements. First test the model to see what it classifies.

- Find how many classes does this model classify. What will appear is that there are 100 classes being classified. Among them is a weird class named “DATASET_DOWNLOAD”. I only provided the last 11 classes.

- Too reach this class, we need to find the array values that trigger this class, but not by brute force. This is done by using
scikit-learnpackage to export the model details, which are the values of each element that triggers a specific class.
- This image shows that the array we needed should have:
- X1 ≤ 10.66
- X2 ≤ 3.23
- X3 ≤ -120.42
- X4, X5, X6 any values, not important
- With this information, you can now go to the server instance and input the array with the intended values.
Phase 2 Solution
After the conditions are met in the server instance, a google drive will be provided, with 2 datasets inside it:
- main.csv
- order.csv
With these 2 datasets, the flag will be found.
Parse
timestampas a true datetime so sorting is chronological rather than lexicographic.The secret column that needs to be focused on is “champanzini_pananini”
Join the two tables on
id. Use the default inner join semantics so only rows present in both files remain—this prevents extra/missing rows from corrupting the hidden bitstream.Sort the merged table by
timestamp. Use a stable sort so rows with identical timestamps keep their original relative order (important for deterministic decoding).Convert the presence/absence of values in
secret_colto bits:- Treat NaN (missing) as 1.
- Treat present (a number) as 0.
This yields a long list of 0/1 bits in the correct temporal order.
Read the bitstream 8 bits at a time. Convert each 8-bit chunk to its integer value, then map that to the corresponding ASCII character.
Final flag is : CONCTF{p01son_WA5_70uNd}
