File size: 13,580 Bytes
df01262 461f097 df01262 52f11c5 df01262 b0db193 df01262 383dc40 1a35659 df01262 383dc40 df01262 b0db193 1a35659 b0db193 df01262 7d9be0d df01262 9cb5aac df01262 52f11c5 df01262 461f097 df01262 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 |
---
license: apache-2.0
datasets:
- b-mc2/sql-create-context
- philikai/Spider-SQL-LLAMA2_train
- ChrisHayduk/Llama-2-SQL-Dataset
language:
- en
metrics:
- accuracy
- f1
- recall
- precision
library_name: keras
pipeline_tag: text-classification
tags:
- sql-injection
- malicious-sql
- sql-injection-detection
- malicious-sql-detection
---
# SafeSQL-v1 ([Playground](https://huggingface.co/spaces/deathsaber93/SafeSQL-v1-Demo))
### Model Meta
- **Feedback:** aakash.howlader@gmail.com
- **Model type:** Language model
- **Language(s) (NLP):** English
- **License:** Apache 2.0
- **Playground:** [SafeSQL-v1-Demo](https://huggingface.co/spaces/deathsaber93/SafeSQL-v1-Demo)
## Overview
This is a Keras 3.x model trained specifically to detect malicious SQLs. It is able to detect various SQL injection vectors such as Error-based, Union-based, Blind, Boolean-based
,Time-based, Out-of-band, Stacked queries. This was trained on ~167K SQLs containing an almost even distribution of malicious and benign SQLs. Its training involved preprocessing specifically for
SQL with special masking tokens. 28 additional numeric features were also generated and the top 10 among them were selected for training using Recursive Feature Elimination. The training consisted of a warm-up period with a smaller, sinusoidally decaying learning rate
followed by a higher learning rate with cosine decay. A special callback was used to monitor for and protect against gradient explosions and
automatically adjust the learning rate and model weights based on the scale of the explosion.
Weight and kernel constraints are applied to help prevent overfitting and achieve better generalization.
For faster model loading and inference, [mixed precision](https://www.tensorflow.org/guide/mixed_precision) has been used.
The best checkpoint has been saved and made available for use.
**CONTEXT WINDOW:** 1200 tokens
**PARAMETERS:** 30.7M *(**Trainable:** 7.7M, **Frozen:** 2K, **Optimizer:** 23M)*
**NUMBER OF INPUTS:** 2 - The SQL queries as string and extra numeric features.
**NUMBER OF OUTPUTS:** 1 - Probability that the given SQL is malicious (the output layer uses a sigmoid activation).
#### Checkpointed Epoch
```
823/823 ━━━━━━━━━━━━━━━━━━━━ 99s 120ms/step - AUPR: 0.9979 - f1_score: 0.5782 - fn: 64.0947 - fp: 8.2500 - loss: 0.0236 - precision: 0.9987 - recall: 0.9889 - val_AUPR: 0.9970 - val_f1_score: 0.5775 - val_fn: 34.0000 - val_fp: 4.0000 - val_loss: 0.0298 - val_precision: 0.9985 - val_recall: 0.9873 - learning_rate: 7.0911e-04
```
#### Benchmark Results
**Total SQLs:** 30919
**Total Negatives:** 11382
**Total Positives:** 19537
**Total hits:** 30844/30919 with accuracy of **99.76%**.
**False Negatives:** 69 - **0.61%**
**False Positives:** 6 - **0.03%**
#### Training Data
The training data is made available [here](dataset/train.csv) and the benchmark data is made
available [here](dataset/benchmark.csv). The data was curated from the following sources -
1. https://www.kaggle.com/datasets/gambleryu/biggest-sql-injection-dataset/data
2. https://huggingface.co/datasets/b-mc2/sql-create-context
3. https://github.com/payloadbox/sql-injection-payload-list/tree/master/Intruder
4. https://huggingface.co/datasets/ChrisHayduk/Llama-2-SQL-Dataset/viewer/default/eval
5. https://huggingface.co/datasets/philikai/Spider-SQL-LLAMA2_train/viewer/default/train
#### Benchmark Data
1. https://www.kaggle.com/datasets/sajid576/sql-injection-dataset?select=Modified_SQL_Dataset.csv
## How to Use
1. Based on your hardware (whether using GPU or not), please download the corresponding `requiremnts-[cpu/gpu].txt` file and install it (`pip install -r requirements.txt`)
2. Download the model file `sqid.keras`.
3. The model expects certain numerical features along with the SQL query. As of v1, some boiler-plate code needs to be written in order to add the numeric features. Please use the below code snippet to load the model, add the expected numeric features and run an inference.
Future iterations of the model will have the numeric features baked into the network's layers.
```Python
import re
from multiprocessing import cpu_count
from keras.src.saving import load_model
import pandas as pd
from numpy import int64
from pandarallel import pandarallel
from sklearn.preprocessing import RobustScaler
model = load_model('./sqid.keras')
pandarallel.initialize(use_memory_fs=True, nb_workers=cpu_count())
def sql_tokenize(sql_query):
sql_query = sql_query.replace('`', ' ').replace('%20', ' ').replace('=', ' = ').replace('((', ' (( ').replace(
'))', ' )) ').replace('(', ' ( ').replace(')', ' ) ').replace('||', ' || ').replace(',', '').replace(
'--', ' -- ').replace(':', ' : ').replace('%23', ' # ').replace('+', ' + ').replace('!=',
' != ') \
.replace('"', ' " ').replace('%26', ' and ').replace('$', ' $ ').replace('%28', ' ( ').replace('%2A', ' * ') \
.replace('%7C', ' | ').replace('&', ' & ').replace(']', ' ] ').replace('[', ' [ ').replace(';',
' ; ').replace(
'/*', ' /* ')
sql_reserved = {'SELECT', 'FROM', 'WHERE', 'AND', 'OR', 'NOT', 'IN', 'LIKE', 'ORDER', 'BY', 'GROUP', 'HAVING',
'LIMIT', 'BETWEEN', 'IS', 'NULL', '%', 'LIKE', 'MIN', 'MAX', 'AS', 'UPPER', 'LOWER', 'TO_DATE',
'=', '>', '<', '>=', '<=', '!=', '<>', 'BETWEEN', 'LIKE', 'EXISTS', 'JOIN', 'UNION', 'ALL',
'ASC', 'DESC', '||', 'AVG', 'LIMIT', 'EXCEPT', 'INTERSECT', 'CASE', 'WHEN', 'THEN', 'IF',
'IF', 'ANY', 'CAST', 'CONVERT', 'COALESCE', 'NULLIF', 'INNER', 'OUTER', 'LEFT', 'RIGHT', 'FULL',
'CROSS', 'OVER', 'PARTITION', 'SUM', 'COUNT', 'WITH', 'INTERVAL', 'WINDOW', 'OVER',
'ROW_NUMBER', 'RANK',
'DENSE_RANK', 'NTILE', 'FIRST_VALUE', 'LAST_VALUE', 'LAG', 'LEAD', 'DISTINCT', 'COMMENT',
'INSERT',
'UPDATE', 'DELETED', 'MERGE', '*', 'generate_series', 'char', 'chr', 'substr', 'lpad',
'extract',
'year', 'month', 'day', 'timestamp', 'number', 'string', 'concat', 'INFORMATION_SCHEMA',
"SQLITE_MASTER", 'TABLES', 'COLUMNS', 'CUBE', 'ROLLUP', 'RECURSIVE', 'FILTER', 'EXCLUDE',
'AUTOINCREMENT', 'WITHOUT', 'ROWID', 'VIRTUAL', 'INDEXED', 'UNINDEXED', 'SERIAL',
'DO', 'RETURNING', 'ILIKE', 'ARRAY', 'ANYARRAY', 'JSONB', 'TSQUERY', 'SEQUENCE',
'SYNONYM', 'CONNECT', 'START', 'LEVEL', 'ROWNUM', 'NOCOPY', 'MINUS', 'AUTO_INCREMENT', 'BINARY',
'ENUM', 'REPLACE', 'SET', 'SHOW', 'DESCRIBE', 'USE', 'EXPLAIN', 'STORED', 'VIRTUAL', 'RLIKE',
'MD5', 'SLEEP', 'BENCHMARK', '@@VERSION', 'VERSION', '@VERSION', 'CONVERT', 'NVARCHAR', '#',
'##', 'INJECTX',
'DELAY', 'WAITFOR', 'RAND',
}
tokens = sql_query.split()
tokens = [re.sub(r"""[^*\w\s.=\-><_|()!"']""", '', token) for token in tokens]
for i, token in enumerate(tokens):
if token.strip().upper() in sql_reserved:
continue
if token.strip().isnumeric():
tokens[i] = '#NUMBER#'
elif re.match(r'^[a-zA-Z_.|][a-zA-Z0-9_.|]*$', token.strip()):
tokens[i] = '#IDENTIFIER#'
elif re.match(r'^[\d:]*$', token.strip()):
tokens[i] = '#TIMESTAMP#'
elif '%' in token.strip():
tokens[i] = ' '.join(
[j.strip() if j.strip() in ('%', "'", "'") else '#IDENTIFIER#' for j in token.strip().split('%')])
return ' '.join(tokens)
def add_features(x):
s = ["num_tables", "num_columns", "num_literals", "num_parentheses", "has_union", "depth_nested_queries", "num_join", "num_sp_chars", "has_mismatched_quotes", "has_tautology"]
x['Query'] = x['Query'].copy().parallel_apply(lambda a: sql_tokenize(a))
x['num_tables'] = x['Query'].str.lower().str.count(r'FROM\s+#IDENTIFIER#', flags=re.I)
x['num_columns'] = x['Query'].str.lower().str.count(r'SELECT\s+#IDENTIFIER#', flags=re.I)
x['num_literals'] = x['Query'].str.lower().str.count("'[^']*'", flags=re.I) + x['Query'].str.lower().str.count(
'"[^"]"', flags=re.I)
x['num_parentheses'] = x['Query'].str.lower().str.count("\\(", flags=re.I) + x['Query'].str.lower().str.count(
'\\)',
flags=re.I)
x['has_union'] = x['Query'].str.lower().str.count(" union |union all", flags=re.I) > 0
x['has_union'] = x['has_union'].astype(int64)
x['depth_nested_queries'] = x['Query'].str.lower().str.count("\\(", flags=re.I)
x['num_join'] = x['Query'].str.lower().str.count(
" join |inner join|outer join|full outer join|full inner join|cross join|left join|right join",
flags=re.I)
x['num_sp_chars'] = x['Query'].parallel_apply(lambda a: len(re.findall(r'[\'";\-*/%=><|#]', a)))
x['has_mismatched_quotes'] = x['Query'].parallel_apply(
lambda sql_query: 1 if re.search(r"'.*[^']$|\".*[^\"]$", sql_query) else 0)
x['has_tautology'] = x['Query'].parallel_apply(lambda sql_query: 1 if re.search(r"'[\s]*=[\s]*'", sql_query) else 0)
return x
input_sqls = ['SELECT roomName , RoomId FROM Rooms WHERE basePrice > 160 AND maxOccupancy > 2;', # Not malicious
"ORDER BY 1,SLEEP(5),BENCHMARK(1000000,MD5('A')),4,5,6,7,8,9,10,11,12,13,14,15,16,17,18#", # Malicious
"; desc users; --", # Malicious
"ORDER BY 1,SLEEP(5),BENCHMARK(1000000,MD5('A')),4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27", # Malicious
"SELECT DISTINCT t2.datasetid FROM paperdataset AS t3 JOIN dataset AS t2 ON t3.datasetid = t2.datasetid JOIN paperkeyphrase AS t1 ON t1.paperid = t3.paperid JOIN keyphrase AS t4 ON t1.keyphraseid = t4.keyphraseid WHERE t4.keyphrasename = ""semantic parsing"";" # Not malicious
]
numeric_features = ["num_tables", "num_columns", "num_literals", "num_parentheses", "has_union", "depth_nested_queries", "num_join", "num_sp_chars", "has_mismatched_quotes", "has_tautology"]
input_df = pd.DataFrame(input_sqls, columns=['Query'])
input_df = add_features(input_df)
scaler = RobustScaler()
x_in = scaler.fit_transform(input_df[numeric_features])
preds = model.predict([input_df['Query'], x_in]).tolist()
for i, pred in enumerate(preds):
print()
print(f'Query: {input_sqls[i]}')
print(f'Malicious? {pred[0] >= 0.50} ({pred[0]})')
print()
# Run the benchmark
input_df = pd.read_csv('benchmark.csv')
hits = 0
data_size = input_df.shape[0]
miss_pos, miss_neg = [], []
total_negs = input_df[input_df['Label'] == 1.0].shape[0]
total_pos = input_df[input_df['Label'] == 0.0].shape[0]
pred_trans = ['Benign', 'Malicious']
false_metrics = {0: 0, 1: 0}
x_in = scaler.transform(input_df[numeric_features])
print('Running benchmark')
preds = model.predict([input_df['Query'], x_in])
miss_q = []
actuals = input_df['Label'].tolist()
for i, pred in enumerate(preds):
pred = int(pred[0] > .95)
if pred == actuals[i]:
hits += 1
else:
false_metrics[int(pred)] += 1
print('Finished benchmark.')
print('printing results.')
acc = round((hits / data_size) * 100, 2)
f_neg = round((false_metrics[0] / total_negs) * 100, 2)
f_pos = round((false_metrics[1] / total_pos) * 100, 2)
print(f'Total data: {data_size}')
print(f'Total Negatives: {total_negs} \t Total Positives: {total_pos}')
print()
print(f'Total hits: {hits}/{data_size} with accuracy of {acc}%.')
print(f'False Negatives: {false_metrics[0]}({f_neg}%) \t False Positives: {false_metrics[1]}({f_pos}%)', false_metrics[0], f_neg, false_metrics[1], f_pos)
```
#### Output
```
2024-06-16 17:34:54.587073: I tensorflow/core/util/port.cc:113] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
...
2024-06-16 17:36:11.762174: I external/local_xla/xla/stream_executor/cuda/cuda_dnn.cc:465] Loaded cuDNN version 8902
1/1 ━━━━━━━━━━━━━━━━━━━━ 14s 14s/step
Query: SELECT roomName , RoomId FROM Rooms WHERE basePrice > 160 AND maxOccupancy > 2;
Malicious? False (7.727547199465334e-05)
Query: ORDER BY 1,SLEEP(5),BENCHMARK(1000000,MD5('A')),4,5,6,7,8,9,10,11,12,13,14,15,16,17,18#
Malicious? True (1.0)
Query: ; desc users; --
Malicious? True (0.9999552965164185)
Query: ORDER BY 1,SLEEP(5),BENCHMARK(1000000,MD5('A')),4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27
Malicious? True (1.0)
Query: SELECT DISTINCT t2.datasetid FROM paperdataset AS t3 JOIN dataset AS t2 ON t3.datasetid = t2.datasetid JOIN paperkeyphrase AS t1 ON t1.paperid = t3.paperid JOIN keyphrase AS t4 ON t1.keyphraseid = t4.keyphraseid WHERE t4.keyphrasename = semantic parsing;
Malicious? False (6.156865989259686e-11)
Running benchmark
967/967 ━━━━━━━━━━━━━━━━━━━━ 37s 37ms/step
Finished benchmark.
printing results.
Total data: 30919
Total Negatives: 11382 Total Positives: 19537
Total hits: 30844/30919 with accuracy of 99.76%.
False Negatives: 69(0.61%) False Positives: 6(0.03%)
```
## Architecture
 |