error_utils.py 17.2 KB
Newer Older
1
#!/usr/bin/env python
Amir MOHAMMADI's avatar
nit  
Amir MOHAMMADI committed
2 3
# Ivana Chingovska <ivana.chingovska@idiap.ch>
# Fri Dec  7 12:33:37 CET 2012
4 5 6 7 8 9 10 11 12
"""Utility functions for computation of EPSC curve and related measurement"""

import bob.measure
import numpy


def calc_pass_rate(threshold, attacks):
    """Calculates the rate of successful spoofing attacks

13 14 15 16 17 18 19 20 21 22 23 24 25
    Parameters
    ----------
    threshold :
      the threshold used for classification
    scores :
      numpy with the scores of the spoofing attacks

    Returns
    -------
    float
      rate of successful spoofing attacks
    """
    return (attacks >= threshold).mean()
26 27 28 29 30 31 32


def weighted_neg_error_rate_criteria(data,
                                     weight,
                                     thres,
                                     beta=0.5,
                                     criteria='eer'):
33 34 35
    """Given the single value for the weight parameter balancing between
    impostors and spoofing attacks and a threshold, calculates the error rates
    and their relationship depending on the criteria (difference in case of
36
    'eer', hter in case of 'min-hter' criteria)
37 38
    Keyword parameters:

39 40 41 42 43
      - data - the development data used to determine the threshold. List on 4
      numpy.arrays containing: negatives (licit), positives (licit),
      negatives (spoof), positivies (spoof)
      - weight - the weight parameter balancing between impostors and spoofing
      attacks
44
      - thres - the given threshold
45 46 47
      - beta - the weight parameter balancing between real accesses and all the
      negative samples (impostors and spoofing attacks). Note that this
      parameter will be overriden and not considered if the selected criteria
48 49
      is 'min-hter'.
      - criteria - 'eer', 'wer' or 'min-hter' criteria for decision threshold
50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68
  """

    licit_neg = data[0]
    licit_pos = data[1]
    spoof_neg = data[2]
    spoof_pos = data[3]  # unpacking the data
    farfrr_licit = bob.measure.farfrr(licit_neg, licit_pos, thres)
    farfrr_spoof = bob.measure.farfrr(spoof_neg, spoof_pos, thres)

    frr = farfrr_licit[1]  # farfrr_spoof[1] should have the same value
    far_i = farfrr_licit[0]
    far_s = farfrr_spoof[0]

    far_w = (1 - weight) * far_i + weight * far_s

    if criteria == 'eer':
        if beta == 0.5:
            return abs(far_w - frr)
        else:
Amir MOHAMMADI's avatar
nit  
Amir MOHAMMADI committed
69
            # return abs(far_w - frr)
70 71
            return abs((1 - beta) * frr - beta * far_w)

72
    elif criteria == 'min-hter':
73 74 75 76 77 78 79 80 81 82 83 84
        return (far_w + frr) / 2

    else:
        return (1 - beta) * frr + beta * far_w


def recursive_thr_search(data,
                         span_min,
                         span_max,
                         weight,
                         beta=0.5,
                         criteria='eer'):
85 86 87 88 89
    """Recursive search for the optimal threshold given a criteria. It
    evaluates the full range of thresholds at 100 points, and computes the one
    which optimizes the threshold. In the next search iteration, it examines
    the region around the point that optimizes the threshold. The procedure
    stops when the search range is smaller then 1e-10.
90 91

  Keyword arguments:
92 93 94
    - data - the development data used to determine the threshold. List on 4
    numpy.arrays containing: negatives (licit), positives (licit), negatives
    (spoof), positivies (spoof)
95 96
    - span_min - the minimum of the search range
    - span_max - the maximum of the search range
97 98 99 100 101
    - weight - the weight parameter balancing between impostors and spoofing
    attacks
    - beta - the weight parameter balancing between real accesses and all the
    negative samples (impostors and spoofing attacks). Note that methods called
    within this function will override this parameter and not considered if the
102
    selected criteria is 'min-hter'.
103
    - criteria - the decision threshold criteria ('eer' for EER, 'wer' for
104
    Minimum WER or 'min-hter' for Minimum HTER criteria).
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
  """

    quit_thr = 1e-10
    steps = 100
    if abs((span_max - span_min) / span_max) < quit_thr:
        return span_max  # or span_min, it doesn't matter
    else:
        step_size = (span_max - span_min) / steps
        thresholds = numpy.array(
            [(i * step_size) + span_min for i in range(steps + 1)])
        weighted_error_rates = numpy.array([
            weighted_neg_error_rate_criteria(data, weight, thr, beta, criteria)
            for thr in thresholds
        ])
        selected_thres = thresholds[numpy.where(
            weighted_error_rates == min(weighted_error_rates)
        )]  # all the thresholds which have minimum weighted error rate
        thr = selected_thres[int(
            selected_thres.size / 2
        )]  # choose the centrally positioned threshold
        return recursive_thr_search(data, thr - step_size, thr + step_size,
                                    weight, beta, criteria)


def weighted_negatives_threshold(licit_neg,
                                 licit_pos,
                                 spoof_neg,
                                 spoof_pos,
                                 weight,
                                 beta=0.5,
                                 criteria='eer'):
136
    """Calculates the threshold for achieving the given criteria between the
Amir MOHAMMADI's avatar
nit  
Amir MOHAMMADI committed
137
    FAR_w and the FRR, given the single value for the weight parameter
138 139 140
    balancing between impostors and spoofing attacks and a single value for the
    parameter beta balancing between the real accesses and the negatives
    (impostors and spoofing attacks)
141 142 143 144 145 146

  Keyword parameters:
    - licit_neg - numpy.array of scores for the negatives (licit scenario)
    - licit_pos - numpy.array of scores for the positives (licit scenario)
    - spoof_neg - numpy.array of scores for the negatives (spoof scenario)
    - spoof_pos - numpy.array of scores for the positives (spoof scenario)
147 148 149 150 151
    - weight - the weight parameter balancing between impostors and spoofing
    attacks
    - beta - the weight parameter balancing between real accesses and all the
    negative samples (impostors and spoofing attacks). Note that methods called
    within this function will override this parameter and not considered if the
152
    selected criteria is 'min-hter'.
153
    - criteria - the decision threshold criteria ('eer' for EER, 'wer' for
154
    Minimum WER or 'min-hter' for Minimum HTER criteria).
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
  """
    span_min = min(
        numpy.append(licit_neg, spoof_neg)
    )  # the min of the span where we will search for the threshold
    span_max = max(
        numpy.append(licit_pos, spoof_pos)
    )  # the max of the span where we will search for the threshold
    data = (licit_neg, licit_pos, spoof_neg,
            spoof_pos)  # pack the data into a single list
    return recursive_thr_search(data, span_min, span_max, weight, beta,
                                criteria)


def epsc_weights(licit_neg, licit_pos, spoof_neg, spoof_pos, points=100):
    """Returns the weights for EPSC

  Keyword arguments:

    - points - number of points to calculate EPSC
  """
    step_size = 1 / float(points)
    weights = numpy.array([(i * step_size) for i in range(points + 1)])
    return weights


def epsc_thresholds(licit_neg,
                    licit_pos,
                    spoof_neg,
                    spoof_pos,
                    points=100,
                    criteria='eer',
                    omega=None,
                    beta=None):
188 189 190 191
    """Calculates the optimal thresholds for EPSC, for a range of the weight
    parameter balancing between impostors and spoofing attacks, and for a range
    of the beta parameter balancing between real accesses and all the negatives
    (impostors and spoofing attacks)
192 193 194 195 196 197 198 199

  Keyword arguments:

    - licit_neg - numpy.array of scores for the negatives (licit scenario)
    - licit_pos - numpy.array of scores for the positives (licit scenario)
    - spoof_neg - numpy.array of scores for the negatives (spoof scenario)
    - spoof_pos - numpy.array of scores for the positives (spoof scenario)
    - points - number of points to calculate EPSC
200
    - criteria - the decision threshold criteria ('eer', 'wer' or 'min-hter')
201 202 203 204
    - omega - the value of the parameter omega, balancing between impostors and
    spoofing attacks. If None, it is going to span the full range [0,1].
    Otherwise, can be set to a fixed value or a list of values.
    - beta - the value of the parameter beta, balancing between real accesses
Amir MOHAMMADI's avatar
nit  
Amir MOHAMMADI committed
205
    and all the negatives (zero-effort impostors and spoofing attacks). If
206 207
    None, it is going to span the full range [0,1]. Otherwise, can be set to a
    fixed value or a list of values.
208 209 210 211

  """
    step_size = 1 / float(points)

Amir MOHAMMADI's avatar
nit  
Amir MOHAMMADI committed
212
    if omega is None:
213 214 215 216 217 218 219
        omega = numpy.array([(i * step_size) for i in range(points + 1)])
    elif not isinstance(omega, list) and not isinstance(
            omega, tuple) and not isinstance(omega, numpy.ndarray):
        omega = numpy.array([omega])
    else:
        omega = numpy.array(omega)

Amir MOHAMMADI's avatar
nit  
Amir MOHAMMADI committed
220
    if beta is None:
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
        beta = numpy.array([(i * step_size) for i in range(points + 1)])
    elif not isinstance(beta, list) and not isinstance(
            beta, tuple) and not isinstance(beta, numpy.ndarray):
        beta = numpy.array([beta])
    else:
        beta = numpy.array(beta)

    thresholds = numpy.ndarray([beta.size, omega.size], 'float64')
    for bindex, b in enumerate(beta):
        thresholds[bindex, :] = numpy.array([
            weighted_negatives_threshold(
                licit_neg,
                licit_pos,
                spoof_neg,
                spoof_pos,
                w,
                b,
                criteria=criteria) for w in omega
        ], 'float64')

    return omega, beta, thresholds


def weighted_err(error_1, error_2, weight):
    """Calculates the weighted error rate between the two input parameters

  Keyword arguments:
248 249
    - error_1 - the first input error rate (FAR for zero effort impostors
    usually)
250 251 252 253 254 255 256 257 258 259 260 261 262
    - error_2 - the second input error rate (SFAR)
    - weight - the given weight
  """
    return (1 - weight) * error_1 + weight * error_2


def error_rates_at_weight(licit_neg,
                          licit_pos,
                          spoof_neg,
                          spoof_pos,
                          omega,
                          threshold,
                          beta=0.5):
263 264 265
    """Calculates several error rates: FRR, FAR (zero-effort impostors), SFAR,
    FAR_w, HTER_w for a given value of w. It returns the calculated threshold
    as a last argument
266 267 268 269 270 271 272 273

  Keyword arguments:

    - licit_neg - numpy.array of scores for the negatives (licit scenario)
    - licit_pos - numpy.array of scores for the positives (licit scenario)
    - spoof_neg - numpy.array of scores for the negatives (spoof scenario)
    - spoof_pos - numpy.array of scores for the positives (spoof scenario)
    - threshold - the given threshold
274 275 276 277 278
    - omega - the omega parameter balancing between impostors and spoofing
    attacks
    - beta - the weight parameter balancing between real accesses and all the

negative samples (impostors and spoofing attacks).
279 280 281 282 283 284 285 286 287
  """

    farfrr_licit = bob.measure.farfrr(
        licit_neg, licit_pos,
        threshold)  # calculate test frr @ threshold (licit scenario)
    farfrr_spoof = bob.measure.farfrr(
        spoof_neg, spoof_pos,
        threshold)  # calculate test frr @ threshold (spoof scenario)

Amir MOHAMMADI's avatar
nit  
Amir MOHAMMADI committed
288 289
    # we can take this value from farfrr_spoof as well, it doesn't matter
    frr = farfrr_licit[1]
290 291 292 293 294 295 296 297 298 299 300 301
    far = farfrr_licit[0]
    sfar = farfrr_spoof[0]

    far_w = weighted_err(far, sfar, omega)
    hter_w = (far_w + frr) / 2
    wer_wb = weighted_err(frr, far_w, beta)

    return (frr, far, sfar, far_w, wer_wb, hter_w, threshold)


def epsc_error_rates(licit_neg, licit_pos, spoof_neg, spoof_pos, thresholds,
                     omega, beta):
Amir MOHAMMADI's avatar
nit  
Amir MOHAMMADI committed
302 303
    """Calculates several error rates: FAR_w and WER_wb for the given weights
    (omega and beta) and thresholds (the thresholds need to be computed first
304
    using the method: epsc_thresholds() before passing to this method)
305

306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331
    Parameters
    ----------
    licit_neg : array_like
        array of scores for the negatives (licit scenario)
    licit_pos : array_like
        array of scores for the positives (licit scenario)
    spoof_neg : array_like
        array of scores for the negatives (spoof scenario)
    spoof_pos : array_like
        array of scores for the positives (spoof scenario)
    thresholds : array_like
        ndarray with threshold values
    omega : array_like
        array of the omega parameter balancing between impostors
        and spoofing attacks
    beta : array_like
        array of the beta parameter balancing between real accesses
        and all negatives (impostors and spoofing attacks)

    Returns
    -------
    far_w_errors: array_like
        FAR_w
    wer_wb_errors: array_like
        WER_wb
    """
332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349

    far_w_errors = numpy.ndarray((beta.size, omega.size), 'float64')
    wer_wb_errors = numpy.ndarray((beta.size, omega.size), 'float64')

    for bindex, b in enumerate(beta):
        errors = [
            error_rates_at_weight(licit_neg, licit_pos, spoof_neg, spoof_pos,
                                  w, thresholds[bindex, windex], b)
            for windex, w in enumerate(omega)
        ]
        far_w_errors[bindex, :] = [errors[i][3] for i in range(len(errors))]
        wer_wb_errors[bindex, :] = [errors[i][4] for i in range(len(errors))]

    return far_w_errors, wer_wb_errors


def all_error_rates(licit_neg, licit_pos, spoof_neg, spoof_pos, thresholds,
                    omega, beta):
Amir MOHAMMADI's avatar
nit  
Amir MOHAMMADI committed
350 351
    """Calculates several error rates: FAR_w and WER_wb for the given weights
    (omega and beta) and thresholds (the thresholds need to be computed first
352
    using the method: epsc_thresholds() before passing to this method)
353

354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379
    Parameters
    ----------
    licit_neg : array_like
        array of scores for the negatives (licit scenario)
    licit_pos : array_like
        array of scores for the positives (licit scenario)
    spoof_neg : array_like
        array of scores for the negatives (spoof scenario)
    spoof_pos : array_like
        array of scores for the positives (spoof scenario)
    thresholds : array_like
        ndarray with threshold values
    omega : array_like
        array of the omega parameter balancing between impostors
        and spoofing attacks
    beta : array_like
        array of the beta parameter balancing between real accesses
        and all negatives (impostors and spoofing attacks)

    Returns
    -------
    far_w_errors: array_like
        FAR_w
    wer_wb_errors: array_like
        WER_wb
    """
380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400

    frr_errors = numpy.ndarray((beta.size, omega.size), 'float64')
    far_errors = numpy.ndarray((beta.size, omega.size), 'float64')
    sfar_errors = numpy.ndarray((beta.size, omega.size), 'float64')
    far_w_errors = numpy.ndarray((beta.size, omega.size), 'float64')
    wer_wb_errors = numpy.ndarray((beta.size, omega.size), 'float64')
    hter_wb_errors = numpy.ndarray((beta.size, omega.size), 'float64')

    for bindex, b in enumerate(beta):
        errors = [
            error_rates_at_weight(licit_neg, licit_pos, spoof_neg, spoof_pos,
                                  w, thresholds[bindex, windex], b)
            for windex, w in enumerate(omega)
        ]
        frr_errors[bindex, :] = [errors[i][0] for i in range(len(errors))]
        far_errors[bindex, :] = [errors[i][1] for i in range(len(errors))]
        sfar_errors[bindex, :] = [errors[i][2] for i in range(len(errors))]
        far_w_errors[bindex, :] = [errors[i][3] for i in range(len(errors))]
        wer_wb_errors[bindex, :] = [errors[i][4] for i in range(len(errors))]
        hter_wb_errors[bindex, :] = [errors[i][5] for i in range(len(errors))]

401 402
    return (frr_errors, far_errors, sfar_errors, far_w_errors, wer_wb_errors,
            hter_wb_errors)
403 404 405 406 407 408 409 410 411 412 413 414 415 416


def calc_aue(licit_neg,
             licit_pos,
             spoof_neg,
             spoof_pos,
             thresholds,
             omega,
             beta,
             l_bound=0,
             h_bound=1,
             var_param='omega'):
    """Calculates AUE of EPSC for the given thresholds and weights

417
    Keyword arguments:
418 419 420 421 422 423 424 425

    - licit_neg - numpy.array of scores for the negatives (licit scenario)
    - licit_pos - numpy.array of scores for the positives (licit scenario)
    - spoof_neg - numpy.array of scores for the negatives (spoof scenario)
    - spoof_pos - numpy.array of scores for the positives (spoof scenario)
    - l_bound - lower bound of integration
    - h_bound - higher bound of integration
    - points - number of points to calculate EPSC
426
    - criteria - the decision threshold criteria ('eer', 'wer' or 'min-hter')
Amir MOHAMMADI's avatar
nit  
Amir MOHAMMADI committed
427
    - var_param - name of the parameter which is varied on the abscissa
428
    ('omega' or 'beta')
429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451
  """

    from scipy import integrate

    if var_param == 'omega':
        errors = all_error_rates(licit_neg, licit_pos, spoof_neg, spoof_pos,
                                 thresholds, omega, beta)
        weights = omega  # setting the weights to the varying parameter
    else:
        errors = all_error_rates(licit_neg, licit_pos, spoof_neg, spoof_pos,
                                 thresholds, omega, beta)
        weights = beta  # setting the weights to the varying parameter

    wer_errors = errors[4].reshape(1, errors[4].size)

    l_ind = numpy.where(weights >= l_bound)[0][0]
    h_ind = numpy.where(weights <= h_bound)[0][-1]
    aue = integrate.cumtrapz(wer_errors, weights)
    aue = numpy.append(
        [0], aue)  # for indexing purposes, aue is cumulative integration
    aue = aue[h_ind] - aue[l_ind]

    return aue