Adversarial attacks on load forecasting model

Author: Avinash Kumar

.gitignore

@@ -0,0 +1,43 @@
+# Node artifact files
+node_modules/
+dist/
+
+# Compiled Java class files
+*.class
+
+# Compiled Python bytecode
+*.py[cod]
+
+# Log files
+*.log
+
+# Package files
+*.jar
+
+# Maven
+target/
+
+# JetBrains IDE
+.idea/
+
+# Unit test reports
+TEST*.xml
+
+# Generated by MacOS
+.DS_Store
+
+# Generated by Windows
+Thumbs.db
+
+# Applications
+*.app
+*.exe
+*.war
+
+# Large media files
+*.mp4
+*.tiff
+*.avi
+*.flv
+*.mov
+*.wmv

README.md

@@ -1 +1,54 @@
-# adversarial-attacks-on-load-forecasting-model
+# Adversarial Attacks on Load Forecasting Model
+
+## Dataset Information
+### Dataset Used
+1. ENTSO-E Dataset (European Network of Transmission System Operators for Electricity) for hourly load data in Switzerland.
+2. DarkSky Dataset for hourly temperature and weather Icon information of 8 major cities in Switzerland.
+
+### Dataset Raw Features
+There are 21 raw features at each timestamp:
+1. **Load**
+2. **8 Cities temperature**
+3. **8 Cities weather Icon Information** - categorical feature tells which weather icon [categories: icon1, icon2, icon3]
+4. **Holiday** - boolean feature tells weather holiday in switzerland on that date
+5. **Month** - categorical feature tells data of which month  [categories: Jan, Feb, ... , Dec]
+6. **Day** - categorical feature tells data of which day [categories: Mon, Tues, Wed, Thrus, Fri, Sat, Sun]
+7. **Hour** - categorical feature tells data of which hour  [categories: 0, 1, 2, ... , 23]
+
+### Dataset Source
+You can use already preprocessed data present in `data` folder with name `actual_dataset.csv` .
+
+**Feature Vector 77 dimensional at each timestamp:** \
+actual_load - 1 feature \
+8 cities temperature - 8 features \
+8 cities weather icon one hot encoding - (8 cities x 3 categories of icon) = 24 features \
+holiday - 1 feature \
+weekday one hot encoding  - 7 features \
+hour one hot encoding  - 24 features \
+month one hot encoding  - 12 features 
+
+**NOTE:** Please ignore `entsoe` feature column in `actual_dataset.csv`. 
+
+
+## Train Load Forecasting Model
+1. change `DATASET_SPLIT_DATE` in `constants.py` according to how you want to split train and test dataset.
+2. Simply Run
+    ```shell
+    python forecasting.py
+    ```
+3. It will save trained model weights in `output/load_forecasting_model_weights.h5`. save `output/loss_epoch_curve.png` and `output/actual_predicted_load.png` images.
+
+
+## Generate Adversarial Datasets
+1. Simply Run:
+    ```shell
+    python adversarial.py
+    ```
+2. It will generate adversarial datasets for various temperature variation in `data` folder. For Ex: `adversarial_dataset_temp_1.csv` means generate adversarial temperature dataset with 1 Fahrenheit change in temperature.
+
+
+## Results
+Check `results.ipynb` file to see various plots like:
+1. Temperature Profile
+2. Load Forecasting Profile
+3. Forecasting MAPE with Temperature Variation

adversarial.py

@@ -0,0 +1,88 @@
+import numpy as np
+import pandas as pd
+import constants as const
+from dataset import load_dataset
+from forecasting import forecasting_model
+from sklearn.preprocessing import StandardScaler
+
+
+# bound generated adversarial temperature features - attacker capabilities
+def check_constraint(x_orig, x_new):
+    for temp_idx in range(1, 9):
+        x_new[temp_idx] = np.clip(x_new[temp_idx], x_orig[temp_idx] - const.BOUND * np.abs(x_orig[temp_idx]),
+                                  x_orig[temp_idx] + const.BOUND * np.abs(x_orig[temp_idx]))
+    return x_new
+
+
+# calculate signed gradient
+def calculate_signed_gradient(X, temp_idx, model):
+    X_plus = X.copy()
+    X_minus = X.copy()
+    X_plus[0, const.SEQ_LENGTH-1, temp_idx] += const.DELTA
+    X_minus[0, const.SEQ_LENGTH-1, temp_idx] -= const.DELTA
+    gradient = model.predict(X_plus, verbose=0) - model.predict(X_minus, verbose=0)
+    return np.sign(gradient)
+
+
+# Black Box Based Gradient Estimation Algorithm to generate hard to detect adversarial dataset
+def gradient_estimation(df, model, temp_variation):
+    X_adversarial = []
+    data = np.array(df, dtype=float)
+    alpha = const.ALPHA * temp_variation
+
+    # loop over all datapoint
+    for i in range(len(data)):
+        epoch = 1
+        if i < const.SEQ_LENGTH - 1:
+            X_adversarial.append(data[i])
+            continue
+
+        # gamma which denotes to increase or decrease load, if 0 then increase load, if 1 then decrease load
+        gamma = np.random.randint(2)
+        X = data[i - const.SEQ_LENGTH + 1:i + 1].reshape(1, const.SEQ_LENGTH, const.FEATURE_DIM)
+
+        # optimize GRAD_NUM_EPOCHS times
+        while epoch <= const.GRAD_NUM_EPOCHS:
+            for temp_idx in range(1, 9):
+                signed_gradient = calculate_signed_gradient(X, temp_idx, model)
+                if gamma == 0:
+                    X[0][const.SEQ_LENGTH - 1][temp_idx] += alpha * signed_gradient
+                else:
+                    X[0][const.SEQ_LENGTH - 1][temp_idx] -= alpha * signed_gradient
+            epoch = epoch + 1
+
+        # bound temperature data - attacker capabilities
+        X[0][const.SEQ_LENGTH - 1] = check_constraint(data[i], X[0][const.SEQ_LENGTH - 1])
+        # store generated adversarial datapoint
+        X_adversarial.append(X[0][const.SEQ_LENGTH - 1])
+
+    return pd.DataFrame(X_adversarial, index=df.index, columns=df.columns)
+
+
+def generate_adversarial_datasets():
+    # load train and test dataset
+    train_df, test_df = load_dataset(const.DATASET_NAME, const.DATASET_SPLIT_DATE)
+
+    # load saved forecasting model
+    model = forecasting_model(const.SEQ_LENGTH, const.FEATURE_DIM, const.FORECAST_HORIZON)
+    model.load_weights(const.MODEL_NAME)
+    print("loaded trained forecasting model ...")
+
+    # preprocess test dataset based on train dataset standard scaler
+    non_categorical_features = ['actual_load', 'bsl_t', 'brn_t', 'zrh_t', 'lug_t', 'lau_t', 'gen_t', 'stg_t', 'luz_t']
+    features_scaler = StandardScaler()
+    features_scaler.fit(train_df[non_categorical_features])
+    test_df[non_categorical_features] = features_scaler.transform(test_df[non_categorical_features])
+
+    # generate adversarial datasets for various variation in temperature (in Fahrenheit)
+    for i in range(1, const.TEMPERATURE_VARIATION + 1):
+        adversarial_df = gradient_estimation(test_df, model, temp_variation=i)
+        # rescale features to actual values using train data standard scaler
+        adversarial_df[non_categorical_features] = features_scaler.inverse_transform(adversarial_df[non_categorical_features])
+        # save generated adversarial dataset
+        adversarial_df.to_csv('data/adversarial_dataset_temp_' + str(i) + ".csv")
+        print(f'Generated adversarial dataset for temperature variation={i}F')
+
+
+if __name__ == "__main__":
+    generate_adversarial_datasets()

adversarial_attacks_on_load_forecasting.pptx

@@ -0,0 +1,19 @@
+# Dataset Constants
+DATASET_NAME = "data/actual_dataset.csv"
+DATASET_SPLIT_DATE = "2017-5-7"
+
+# Forecasting Model Hyperparameters
+SEQ_LENGTH = 24
+FEATURE_DIM = 77
+FORECAST_HORIZON = 1
+TRAIN_NUM_EPOCHS = 30
+BATCH_SIZE = 32
+
+# Gradient Estimation Hyperparameters
+MODEL_NAME = 'output/load_forecasting_model_weights.h5'
+ALPHA = 0.01
+GRAD_NUM_EPOCHS = 10
+BETA = 0.9
+DELTA = 0.05
+BOUND = 0.4
+TEMPERATURE_VARIATION = 5