Project Sekai - ✅-misc-fighter_jet

Project Sekai

🔒 CrewCTF 2023 / ✅-misc-fighter_jet

Fighter_jet - 1000 points

Category: Misc Description: Come check out our new plane, the THC-16! It's got an awesome state of the art radar and radar warning receiver that gives tons of info! We just haven't gotten around to building the part that intereprets the data yet. Maybe you can help us on your test flight? We've got some test data we can give you, but after that you're on your own. You'll need to determine whether new radar contacts are friendly or enemy. If you can get 90% right then we'll call that a success! For the radar warning receiver you'll need to distinguish between a ground radar locking you and a missile launch. You'll need to launch countermeasures if it's a missile launch. There is no margin for error here. Author: LinuxBro nc fighterjet.chal.crewc.tf 40001 Files: No files. Tags: No tags.

Sutx pinned a message to this channel. 07/07/2023 10:02 PM

@enscribe wants to collaborate

@Surg wants to collaborate

there is a bunch of lock/launch after the data

01:25

ok so there is radio data and rwr data

01:27

radio data 10 dim, rwr 3 dim, weird.. i was assuming radio data is the actual data, then rwr data are train set

@Violin wants to collaborate

@Legoclones wants to collaborate

sahuang

radio data 10 dim, rwr 3 dim, weird.. i was assuming radio data is the actual data, then rwr data are train set

kinda too guessy, basically 2 sets of binary data, first set has dim 10 second dim3, and we need to build the classifier based on first, then train second and predict inputs but i have little idea why their dim doesnt match

wait is this the ML one?

yeah

shouldnt be hard, i just cant work on it rn

06:38

3 solves already

idk ML

msfrog

will u be able to solve at some point?

idts

06:39

also idk ML too, just find random docs and try

06:39

in fact this may or may not be ML depending on the data given

let me read more about the description

06:52

i think i still didnt fully understand the info

06:54

DUMPING PRIOR RADAR DATA
6.367055039289703,0.14929640237488623,-0.1804751685355639,9.063589897217922,-8.462679767550664,-9.522784378997551,-8.534154068111103,7.405369299400735,5.68800574515817,6.253355533975407,FRIENDLY
0.9220112982712879,4.55406027061954,1.4766102234152765,1.0327673642434532,-0.09791864185107468,2.3389504197232576,-2.2684893178207632,8.732543694179428,9.139365673010275,-2.2447658390116523,BANDIT
......
DUMPING PRIOR RWR DATA
-0.10528410529750024,0.7137860299568659,0.0,MISSILE LAUNCH
0.9933574488386961,0.005606370727041129,0.0,GROUND RADAR LOCK
-0.5002389359959747,-0.5198484352554837,0.0,MISSILE LAUNCH
......
SPIKE!  -0.8379939128039027,0.19957742935917555,0.5328467496143284
Is this an incoming missile? [Y/N]:

06:54

whats friendly/bandit vs ground radar lock/missile launch?

06:55

looking at the coordinates they have some relation

06:55

@Legoclones bandit means missile launch?

06:55

in english context

06:56

also idk whats ground radar loxk

06:57

i have a few guesses but idk which is right cuz i cant calculate rn

friendly is good, bandit is bad

06:59

"Bandit" is an identified unfriendly aircraft

radar lock when good?

07:00

missile launch when bad?

07:01

oh lmao

07:01

they have 2 types of queries

Image attachment

07:01

so those seem 2 separate datasets

07:02

missile and friendly or not

07:02

i see

07:02

this seems very doable lol

07:04

just build 2 separate binary supervised model then predict

okay I'm super confused what the data means

There are 2 datasets, the first have entries of 10 dimensions and a label, the second is 3 dimension. Both datasets have 2 labels only, so it is a binary classification

07:05

You dont need to know what each float value means

so how can we use labels to predict missile from spike?

After training 2 models based on each of the dataset, predict based on input

ohh

if input is dimension 3, use the second model

07:06

if 10, use first

so input numbers to model, model will make sense of it and determine if it's good or bad based on previous data

based on the given dataset

missile launch bad, bandit bad

those 2 are unrelated

07:07

they give 2 separate datasets

yes

each is binary classifier

but if they give 10 numbers, use first dataset, and bandit means bad

yes

if they give 3 numbers, use second dataset and missile launch means bad

their question is, is it friendly for dim 10 (edited)

07:07

second question is whether missile launch

how would i even code this lol

https://towardsdatascience.com/building-classification-models-with-sklearn-6a8fd107f0c1

Building Classification Models with Sklearn

Introduction to Classification with Sklearn

07:09

use same code to train and predict

07:09

its simple actually

07:09

just import, add data, train, predict

alright I will try it

07:10

this is good intro to some ml msfrog

msfrog

if you can follow and make a template code ill also check later

cool

okay looks like nearest neighbor classification system would work

07:19

gotta retrieve data, parse into pandas df, use as training data, then output to nc session

idk if knn works, cuz binary classifier

07:41

knn is unsupervised, may work if dataset is good

07:41

otherwise can try logistic in that blog

okay was able to turn incoming data into data frame and capture query

07:48

now need to train model

07:49

I'll try the link u sent me

07:50

What I have so far

1.39 KB

Image attachment

Legoclones

okay was able to turn incoming data into data frame and capture query

yeah

lol it says "train the model" but no code

Image attachment

08:00

wait maybe code at the bottom

08:00

wtf

check doc

08:02

.train or sth

Okay here is updated with linear regression model. I temporarily split it 75% train, 25% test then printed out how good it is. Not very good lol

2.26 KB

08:26

Image attachment

okay actually I think it does work right

08:37

I'm guessing data is randomly generated, which is why splitting test data into test and train is bad.

08:37

let me finish scripting this

Legoclones

I'm guessing data is randomly generated, which is why splitting test data into test and train is bad.

oh maybe?

08:43

then what you gonna use, nearest neighbour?

sticking with logical regression for now, if it doesn't work i might switch over

08:43

logical regression has given good results so far

ic

hmm now it's almost always wrong

almost always?

08:53

reverse the submission result then

we can sanity check w admin if logistic regression makes sense, not too sure

08:53

or knn

Smth smth https://xkcd.com/2270/

xkcd: Picking Bad Stocks

1

successfully got through only 9

Image attachment

09:01

i dont think it's working right, lemme try a bit more, and switch over to knn

09:05

knn not much better

09:05

I'll message admin and send script in here if anyone has any suggestions

09:05

Currently, SPIKE means 3 points in query, and I have this doing KNN. BOGEY means 10 points in query, and I have this doing linear regression

3.34 KB

they said 90% correct, why it exits when one is wrong?

idk

09:21

finally figured out how modmail works lol

09:21

that is jank (edited)

"You'll need to determine whether new radar contacts are friendly or enemy. If you can get 90% right then we'll call that a success! For the radar warning receiver you'll need to distinguish between a ground radar locking you and a missile launch. You'll need to launch countermeasures if it's a missile launch. There is no margin for error here. "

09:21

you can have 90% acc on friends/enemy

09:21

missile is 100% acc

missile is the 3 dim one?

yeah

so maybe they have two different algos

09:22

missile one maybe knn?

09:23

nearest neighbour or whatever

09:23

that is 100% correct, logictic regression has chances of failure

09:23

thats just a guess

09:24

i wonder why they split to two different datasets, possible explanation is two models

knn and regression both aren't very accurate for missiles

09:24

I seem to be doing okay with bogeys, but missiles is the big problem

yeah 3 dimensions

09:25

can you plot the xyz coords?

09:25

and observe if theres any rule

got any shortcuts for plotting dataframes?

just matplotlib on xyz

09:28

https://jakevdp.github.io/PythonDataScienceHandbook/04.12-three-dimensional-plotting.html

Three-Dimensional Plotting in Matplotlib | Python Data Science Hand...

yoooo that's kinda dope

09:36

this is what it looks like from one side

Image attachment

09:36

kind of like a football or rugby ball

09:37

now gotta color code based on bogey or not msfrog

msfrog

09:40

okay yeah inside vs outside

Image attachment

neat

lol mod "the solution works here so i guess keep trying"

09:52

finally figured out how to add query in there

Image attachment

09:52

now just gotta label, then maybe manually will work

welp

10:09

even looking at it, it's still hard to know if it's good or bad

10:09

I only get it right 50% of the time lol

10:09

And now I keep getting bogey ones wrong....

10:09

Idk what I'm doing wrong

try dim reduction?

3d -> 2d?

that or 10d -> 2d

10:10

PCA does wonders sometimes

oh wait

10:11

i'm just dum hold up

Legoclones

finally figured out how to add query in there

is there any rule

yeah lemme illustrate

11:32

it's not super easy to tell, but this is what I have so far. In 3d mode, it's always pretty hard to see, but if you put in this orientation, its clearer. If it's inside the hexagon, it's the light blue. If it's outside, it's purple. Legend will tell on the right which color is good.

Image attachment

ah

okay it's just so hard to tell manually

11:58

okay tbh don't know where to go from here either, I'll keep trying but idk why binary classification doesn't work, and it's so hard to tell manually

leave your scripts, i will check tonight if free

4.14 KB

12:05

I'm having some real skill issues today

12:05

feel like I can't solve anything today

I've been trying different things, like which quadrant it's in, how close it is to any other dot, how close it is to a "good" dot, etc., and still don't always get it right

12:25

Going by distance from a "good" dot has gotten me closest, can do like 4 in a row, but even that's only 80% accurate, and idk how many tests they have

is the missile one 100% win now?

12:26

by plot or scripting?

no, the missile one is 80%

12:26

by plot

oh

12:26

so plot doesnt work?

the scripting non-missile one works great

oh

12:27

10 dim one works 90%?

around 90% yeah, closer to 100% lol

kk

if its an inside/outside correlation

12:27

maybe going to 2D will net better scripting results

12:28

idk how much eyeballing you can pull off, and it might just be 1000 tests

its 3d

12:28

real world 3d missile

sure..., but say we take every data point to be a line segment from origin

12:30

visually, it looks like the first 90% of that segment is not a missle, where the last 10% is (numbers are arbitrary here)

12:31

but the direction of that segment doesn't seem to have any correlation

12:31

so, in theory, thats a rendudant dimension

alright admin said 30 tests

12:37

lol how does this make sense. Them: "Is it a missile?" Me: "Yes, it's a missile. Take countermeasures" Them: "Okay. Oh no! A missile hit us!"

Image attachment

okay I wonder if something is wrong with my script

12:48

The "DISPENSING COUNTERMEASURES" should only ever appear when I type Y

12:48

and yet

Image attachment

12:52

yeah maybe that's why

Image attachment

Hmm

sahuang

reverse the submission result then

Did you follow this advice and forget to change it :p

I literally sent N and it thought it was Y...

12:53

this is the code

Image attachment

12:54

Anything that's not 'N'is interpreted as yes

12:54

So I always do uppercase

12:57

BRO ITS THE NEWLINE

12:58

I SWEAR I CHECKED FOR THAT

I thought that

12:58

But I also thought input strips it

12:58

(Which is why I deleted it)

sameee

Just check first character

12:58

To be N

12:58

Anything else is yes

I just did 'N' in rec

12:59

regardless of location lol

13:02

okay I think I got this now

13:02

I think it depends on which it's closer to

13:02

just got 6 in a row

13:02

now that I know it's actually sending the right results smh

Legoclones

used /ctf solve

✅ Challenge solved.

golly that took WAYY too long

13:07

if closer to purple dots, then GOOD (send N). if closer to green dots, then BAD (send Y)

13:07

got right 10x in a row

13:07

finally did something lol

Nice!

lmaoo

13:29

nice

Exported 214 message(s)