"According to my friend Zuck, the first step on the path to great power is to
rate the relative hotness of stuff... think Hot or Not."
Setup
-----
We are provided a JPEG image.
hotornot.jpg shows the original in its entirety, with the caviot that this one
is scaled WAY down and compressed more heavily. The actual original is very
large (19488x19488, 69 MB).
hotornot_detail.jpg shows a small detailed portion of the original image. We
can see that the original is made up of a grid of images, either pictures of
hotdogs or pictures of dogs.
Ideas
-----
This was a steganography task. As with tasks like this we first try the usual
cheap searching techniques, strings, binwalk, and hexdump to look for anything
that sticks out. Once reasonably sure there was nothing there, we moved on to
the image data.
The original image, as stated before, is a grid of embedded images. Each one
measured to be 224x224 pixels. The entire image is 19488x19488, which means it
is 87 embedded images across and 87 tall.
Given the state of things and the clues dropped in the task title and
description, we were fairly confident in guessing this was a game of
hotdog-notdog.
This comes from an episode of the HBO show "Silicon Valley", where one of the
characters designs a mobile app which can identify food. In reality, it can
only identify hotdogs and will otherwise report that the picture is *not* of a
hotdog. Today I learned that this is actually a real thing, with a trained AI
and public API on the internet. As I'm going back and writing this up, I can't
find the website we used anymore. But here's the API from my script
"https://api.deepai.org/hot_dog_or_not". Hit that with a POST to have an image
analyzed (as we will see below).
With a plan formulating, we are theorizing on what the hotdog-notdog results
will mean. We want to try interpreting the results as bits, packing every 8
into a byte, and reading the resulting stream. Otherwise, we will try
overlaying the results back on the original image to see if a visual pattern
emerges.
Solution
--------
# Image Analysis
Starting with our original image above, we want to offload the analysis of each
embedded image to the API we found. As a convention, I will be refering to
these embedded images as cells.
The first step was extracting each cell into its own image file. For this I
wrote a C program utilizing libjpeg (See Appendix A). This left us with a
directory of files "imgXxY.jpg", where X and Y are the corresponding row and
column of the cell in grid-space [0, 87).
The website we found the AI engine on had some minimal documentation on its API
and left us with a curl command to interact with it. It just takes a POST
passing it the URL to an image to analyze.
curl -X POST -d '{"img_url": "http://..."}' https://api.deepai.org/hot_dog_or_not
Note that it takes images by URL when using the API. So we had to host the
images somewhere. I just resorted to our trusty internal server and opened the
firewall gates!
To automate the process of calling the API with each image, the following
scripts were written. The first Python script just procedurally generates all
the filenames of all the cell images. The second Bash script takes each name,
plugs it into a full URL and makes the curl call, writing the results to a
image-specific text file.
#!/usr/bin/env python
for row in range(87):
for col in range(87):
print(f"img{row}x{col}.jpg")
#!/bin/bash -e
# Read names off stdin, tell deepai to check it out from host #
while read name; do
URL="{\"img_url\":\"http://<REDACTED>/cells/$name\"}"
curl -X POST -d $URL 'https://api.deepai.org/hot_dog_or_not' >results/$name.txt
done
We now have a directory of imgXxY.jpg files and another of imgXxY.jpg.txt files.
Each text file contains the analysis result of its corresponding cell image file.
The contents of the text files looks like this:
{"is_hot_dog": "hot dog"}
in the case of a hotdog, and
{"is_hot_dog": "not hot dog"}
in the case of a notdog. So, we can check one by searching it for the word
"not".
# Interpreting Data
We first tried inspecting the bitstream which resulted from image analysis.
Calling each hotdog a 1 and each notdog a 0 and packing each 8 into a byte, no
recognizable data can be found in the resulting bitstream. The same happens
when calling hotdogs 0 and vice-virsa. We know that there are some false
results back from the AI, but were assuming that the data would be more-or-less
kinda readable as a starting point.
Moving on to an image overlay. We produce a copy of the original image, with
each hotdog covered with white pixels and each notdog covered with black pixels
(see Appendix B). No patterns initially jumped out, but after a while of
studying I started to think I was seeing a QR code in the noise. We both
confirmed some characteristics of the image to gauge the likelyhood of this and
decided to presue it.
## QR Code
'hotornot_initial_overlay.jpg' shows our initial image overlay. The cell
resolution is still 87x87, which is not a valid code according to the standard.
We noticed that the original image happened to always contain hotdog and notdog
images in clusters of 3x3 cells, which was really useful for two reasons. First,
this gave us a way to weed-out some false results from the AI, just group cells
together and produce a cluster based on the majority result. Second, once we
scaled-down our image by replacing cells with clusters, we have an image which
is 29x29 clusters, and this *is* a valid code version as described in the
standard. It's version 3.
We regenerated our QR code using a cluster-oriented approach (See Appendix C).
We also inverted the colors since we noticed a few things that didn't quite make
sense in the original. This produced 'hotornot_refined_overlay.jpg'. This is
still not a valid QR code as it is missing the large markers in the corners and
a few pixels are still malformed according to the version 3 standard. In
particular, this included some alternating patterns along the top and left-hand
side, as well as the small square marker towards the bottom-right. We suspect
these subtle remaining flaws were unintended by the problem designers, since we
can see what images we expect back in the original. It was probably just
unlucky bad results from the AI.
We fixed up the small errors and manually drew in the large corner markers.
This produced 'hotornot_final_overlay.png'. Upon scanning this image with a QR
code reader, we finally recovered the flag.
IceCTF{h0td1gg1tyd0g}
== Appendices ==
=== Appendix A ===
C code to split original image into cells
#include <stdio.h>
#include <stdlib.h>
#include <jpeglib.h>
void write_cell(unsigned int matr, unsigned int matc,
unsigned char *buff, unsigned long fwid)
{
struct jpeg_compress_struct cinfo;
struct jpeg_error_mgr jerr;
/* this is a pointer to one row of image data */
JSAMPROW row_pointer[1];
char fname[32];
sprintf(fname, "img%ix%i.jpg", matr, matc);
FILE *outfile = fopen(fname, "wb"); // res
if (!outfile)
{
fprintf(stderr, "Failed to open output file '%s'\n", fname);
return;
}
cinfo.err = jpeg_std_error( &jerr );
jpeg_create_compress(&cinfo);
jpeg_stdio_dest(&cinfo, outfile);
/* Setting the parameters of the output file here */
cinfo.image_width = 224;
cinfo.image_height = 224;
cinfo.input_components = 3;
cinfo.in_color_space = JCS_RGB;
/* default compression parameters, we shouldn't be worried about these */
jpeg_set_defaults( &cinfo );
/* Now do the compression .. */
jpeg_start_compress( &cinfo, TRUE );
unsigned long row = matr * 224;
unsigned long col = matc * 224;
while (cinfo.next_scanline < cinfo.image_height)
{
unsigned long idx = (cinfo.next_scanline + row) * fwid * cinfo.input_components;
idx += col * cinfo.input_components;
row_pointer[0] = &buff[idx];
jpeg_write_scanlines(&cinfo, row_pointer, 1);
}
/* similar to read file, clean up after we're done compressing */
jpeg_finish_compress( &cinfo );
jpeg_destroy_compress( &cinfo );
fclose( outfile );
}
int main(int argc, char **argv)
{
if (argc < 2)
{
fprintf(stderr, "Usage: %s <file>\n", argv[0]);
return 1;
}
char *file = argv[1];
struct jpeg_decompress_struct jds;
struct jpeg_error_mgr err;
FILE *f;
f = fopen(file, "rb"); // res
if (!f)
{
fprintf(stderr, "Failed to open file\n");
return 1;
}
jds.err = jpeg_std_error(&err);
jpeg_create_decompress(&jds); // res
jpeg_stdio_src(&jds, f);
jpeg_read_header(&jds, TRUE);
jpeg_start_decompress(&jds);
unsigned long width = jds.output_width;
unsigned long height = jds.output_height;
unsigned char *buff = malloc(width * height * 3); // res
unsigned char *tmp[1];
if (!buff)
{
fprintf(stderr, "Could not allocate image buffer\n");
jpeg_finish_decompress(&jds);
return 1;
}
while (jds.output_scanline < height)
{
tmp[0] = buff + (3 * width * jds.output_scanline);
jpeg_read_scanlines(&jds, tmp, 1);
}
jpeg_finish_decompress(&jds);
jpeg_destroy_decompress(&jds);
fclose(f);
/* we have the img in memory now, write every 224x224
* block of pixels out to a matrix of files 'img0x0.jpg' */
unsigned int matr = 0;
unsigned int matc = 0;
printf("width: %li\n", width);
printf("height: %li\n", height);
/* lines */
for (matr = 0; (matr * 224) < height; matr++)
{
/* cells */
for (matc = 0; (matc * 224) < width; matc++)
write_cell(matr, matc, buff, width);
}
printf("rows: %i\n", matr);
printf("cols: %i\n", matc);
free(buff);
return 0;
}
=== Appendix B ===
C code to produce initial QR code image
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <jpeglib.h>
void interpret_cell(unsigned int matr, unsigned int matc,
unsigned char *buff, unsigned long fwid)
{
char results_file[32], line[64];
sprintf(results_file, "deepai/results/img%ix%i.jpg.txt", matr, matc);
FILE *rf = fopen(results_file, "r"); // res
if (!rf)
{
fprintf(stderr, "Failed to open results file '%s'\n", results_file);
return;
}
fgets(line, 64, rf);
char *not = strstr(line, "not");
unsigned long height = 224;
unsigned long width = 224;
unsigned char *linebuff = malloc(3 * width); // res
if (!linebuff)
{
fprintf(stderr, "Failed to allocate result tmp buffer\n");
fclose(rf);
return;
}
if (not)
memset(linebuff, 0, 3*width);
else
memset(linebuff, 255, 3*width);
unsigned long i;
unsigned long row = matr * 224;
unsigned long col = matc * 224;
for (i = 0; i < height; i++)
{
unsigned long idx = (i + row) * fwid * 3;
idx += col * 3;
unsigned char *ptr = &buff[idx];
memcpy(ptr, linebuff, 3*width);
}
free(linebuff);
fclose(rf);
}
void write_img(unsigned char *buff, unsigned long width, unsigned long height)
{
struct jpeg_compress_struct cinfo;
struct jpeg_error_mgr jerr;
/* this is a pointer to one row of image data */
JSAMPROW row_pointer[1];
char fname[32];
sprintf(fname, "myoutput.jpg");
FILE *outfile = fopen(fname, "wb"); // res
if (!outfile)
{
fprintf(stderr, "Failed to open output file '%s'\n", fname);
return;
}
cinfo.err = jpeg_std_error( &jerr );
jpeg_create_compress(&cinfo);
jpeg_stdio_dest(&cinfo, outfile);
/* Setting the parameters of the output file here */
cinfo.image_width = width;
cinfo.image_height = height;
cinfo.input_components = 3;
cinfo.in_color_space = JCS_RGB;
/* default compression parameters, we shouldn't be worried about these */
jpeg_set_defaults( &cinfo );
/* Now do the compression .. */
jpeg_start_compress( &cinfo, TRUE );
while (cinfo.next_scanline < cinfo.image_height)
{
row_pointer[0] = &buff[cinfo.next_scanline * width * cinfo.input_components];
jpeg_write_scanlines(&cinfo, row_pointer, 1);
}
/* similar to read file, clean up after we're done compressing */
jpeg_finish_compress( &cinfo );
jpeg_destroy_compress( &cinfo );
fclose( outfile );
}
int main(int argc, char **argv)
{
if (argc < 2)
{
fprintf(stderr, "Usage: %s <file>\n", argv[0]);
return 1;
}
char *file = argv[1];
struct jpeg_decompress_struct jds;
struct jpeg_error_mgr err;
FILE *f;
f = fopen(file, "rb"); // res
if (!f)
{
fprintf(stderr, "Failed to open file\n");
return 1;
}
jds.err = jpeg_std_error(&err);
jpeg_create_decompress(&jds); // res
jpeg_stdio_src(&jds, f);
jpeg_read_header(&jds, TRUE);
jpeg_start_decompress(&jds);
unsigned long width = jds.output_width;
unsigned long height = jds.output_height;
unsigned char *buff = malloc(width * height * 3); // res
unsigned char *tmp[1];
if (!buff)
{
fprintf(stderr, "Could not allocate image buffer\n");
jpeg_finish_decompress(&jds);
return 1;
}
while (jds.output_scanline < height)
{
tmp[0] = buff + (3 * width * jds.output_scanline);
jpeg_read_scanlines(&jds, tmp, 1);
}
jpeg_finish_decompress(&jds);
jpeg_destroy_decompress(&jds);
fclose(f);
unsigned int matr = 0;
unsigned int matc = 0;
printf("width: %li\n", width);
printf("height: %li\n", height);
/* lines */
for (matr = 0; (matr * 224) < height; matr++)
{
/* cells */
for (matc = 0; (matc * 224) < width; matc++)
interpret_cell(matr, matc, buff, width);
}
printf("rows: %i\n", matr);
printf("cols: %i\n", matc);
write_img(buff, width, height);
free(buff);
return 0;
}
=== Appendix C ===
C code to produce clustered QR code image. We switched to PNG format to
eliminate JPEG compression artifacts. This program also writes an image that is
1 pixel per cluster, to aid manual editing.
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#define PNG_DEBUG 3
#include <png.h>
#define OUT_DIM 29
#define IMG_LEN ((OUT_DIM*OUT_DIM)/8)
#define ORIG_DIM 87
#define FILE_NAME "myoutput.png"
unsigned int get_px_idx(unsigned int wid, unsigned int row,
unsigned int col)
{
return (row*wid) + col;
}
/* transform from cluster-space into cell-space */
unsigned int get_cell_output_px_idx(unsigned cell_row,
unsigned cell_col)
{
return get_px_idx(OUT_DIM, cell_row/3, cell_col/3);
}
void write_image(unsigned char *buff)
{
png_structp png_ptr;
png_infop info_ptr;
/* create file */
FILE *fp = fopen(FILE_NAME, "wb");
if (!fp)
printf("[write_png_file] File %s could not be opened for writing", FILE_NAME);
/* initialize stuff */
png_ptr = png_create_write_struct(PNG_LIBPNG_VER_STRING, NULL, NULL, NULL);
if (!png_ptr)
printf("[write_png_file] png_create_write_struct failed");
info_ptr = png_create_info_struct(png_ptr);
if (!info_ptr)
printf("[write_png_file] png_create_info_struct failed");
if (setjmp(png_jmpbuf(png_ptr)))
printf("[write_png_file] Error during init_io");
png_init_io(png_ptr, fp);
/* write header */
if (setjmp(png_jmpbuf(png_ptr)))
printf("[write_png_file] Error during writing header");
png_set_IHDR(png_ptr, info_ptr, OUT_DIM, OUT_DIM,
8, PNG_COLOR_TYPE_GRAY, PNG_INTERLACE_NONE,
PNG_COMPRESSION_TYPE_DEFAULT, PNG_FILTER_TYPE_DEFAULT);
png_write_info(png_ptr, info_ptr);
/* write bytes */
if (setjmp(png_jmpbuf(png_ptr)))
printf("[write_png_file] Error during writing bytes");
unsigned int row;
for (row=0; row < OUT_DIM; row++)
png_write_row(png_ptr, &buff[get_px_idx(OUT_DIM, row, 0)]);
//png_write_image(png_ptr, &buff);
/* end write */
if (setjmp(png_jmpbuf(png_ptr)))
printf("[write_png_file] Error during end of write");
png_write_end(png_ptr, NULL);
fclose(fp);
}
int interpret_cell(unsigned int cell_row, unsigned int cell_col)
{
char fname[32], line[64];
snprintf(fname, 32, "deepai/results/img%ix%i.jpg.txt",
cell_row, cell_col);
FILE *f = fopen(fname, "r");
if (!f)
{
fprintf(stderr, "Failed to open result file '%s'\n", fname);
return 0;
}
fgets(line, 64, f);
char *not = strstr(line, "not");
fclose(f);
return not == NULL;
}
int main()
{
/* output image information */
unsigned char *outbuff = malloc(OUT_DIM*OUT_DIM);
memset(outbuff, 0, OUT_DIM*OUT_DIM);
/* iterate original cells (not clusters) */
unsigned int cell_row, cell_col;
for (cell_row = 0; cell_row < ORIG_DIM; cell_row++)
{
for (cell_col = 0; cell_col < ORIG_DIM; cell_col++)
{
int res = interpret_cell(cell_row, cell_col);
outbuff[get_cell_output_px_idx(cell_row, cell_col)] += res;
}
}
/* counts to colors */
unsigned int r, c;
for (r = 0 ; r < OUT_DIM; r++)
{
for (c = 0; c < OUT_DIM; c++)
{
unsigned int idx = get_px_idx(OUT_DIM, r, c);
if (outbuff[idx] >= 5)
outbuff[idx] = 0;
else
outbuff[idx] = 255;
}
}
/* write image */
write_image(outbuff);
free(outbuff);
return 0;
}
